1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86InstrBuilder.h"
19 #include "X86MachineFunctionInfo.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/StringSwitch.h"
26 #include "llvm/ADT/VariadicFunction.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineJumpTableInfo.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/IR/CallSite.h"
35 #include "llvm/IR/CallingConv.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DerivedTypes.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/GlobalAlias.h"
40 #include "llvm/IR/GlobalVariable.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/MC/MCAsmInfo.h"
44 #include "llvm/MC/MCContext.h"
45 #include "llvm/MC/MCExpr.h"
46 #include "llvm/MC/MCSymbol.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Target/TargetOptions.h"
57 #define DEBUG_TYPE "x86-isel"
59 STATISTIC(NumTailCalls, "Number of tail calls");
61 static cl::opt<bool> ExperimentalVectorWideningLegalization(
62 "x86-experimental-vector-widening-legalization", cl::init(false),
63 cl::desc("Enable an experimental vector type legalization through widening "
64 "rather than promotion."),
67 static cl::opt<bool> ExperimentalVectorShuffleLowering(
68 "x86-experimental-vector-shuffle-lowering", cl::init(false),
69 cl::desc("Enable an experimental vector shuffle lowering code path."),
72 // Forward declarations.
73 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
76 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
77 SelectionDAG &DAG, SDLoc dl,
78 unsigned vectorWidth) {
79 assert((vectorWidth == 128 || vectorWidth == 256) &&
80 "Unsupported vector width");
81 EVT VT = Vec.getValueType();
82 EVT ElVT = VT.getVectorElementType();
83 unsigned Factor = VT.getSizeInBits()/vectorWidth;
84 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
85 VT.getVectorNumElements()/Factor);
87 // Extract from UNDEF is UNDEF.
88 if (Vec.getOpcode() == ISD::UNDEF)
89 return DAG.getUNDEF(ResultVT);
91 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
92 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
94 // This is the index of the first element of the vectorWidth-bit chunk
96 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
99 // If the input is a buildvector just emit a smaller one.
100 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
101 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
102 makeArrayRef(Vec->op_begin()+NormalizedIdxVal,
105 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
106 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
112 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
113 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
114 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
115 /// instructions or a simple subregister reference. Idx is an index in the
116 /// 128 bits we want. It need not be aligned to a 128-bit bounday. That makes
117 /// lowering EXTRACT_VECTOR_ELT operations easier.
118 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
119 SelectionDAG &DAG, SDLoc dl) {
120 assert((Vec.getValueType().is256BitVector() ||
121 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
122 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
125 /// Generate a DAG to grab 256-bits from a 512-bit vector.
126 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
127 SelectionDAG &DAG, SDLoc dl) {
128 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
129 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
132 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
133 unsigned IdxVal, SelectionDAG &DAG,
134 SDLoc dl, unsigned vectorWidth) {
135 assert((vectorWidth == 128 || vectorWidth == 256) &&
136 "Unsupported vector width");
137 // Inserting UNDEF is Result
138 if (Vec.getOpcode() == ISD::UNDEF)
140 EVT VT = Vec.getValueType();
141 EVT ElVT = VT.getVectorElementType();
142 EVT ResultVT = Result.getValueType();
144 // Insert the relevant vectorWidth bits.
145 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
147 // This is the index of the first element of the vectorWidth-bit chunk
149 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
152 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
153 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
156 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
157 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
158 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
159 /// simple superregister reference. Idx is an index in the 128 bits
160 /// we want. It need not be aligned to a 128-bit bounday. That makes
161 /// lowering INSERT_VECTOR_ELT operations easier.
162 static SDValue Insert128BitVector(SDValue Result, SDValue Vec,
163 unsigned IdxVal, SelectionDAG &DAG,
165 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
166 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
169 static SDValue Insert256BitVector(SDValue Result, SDValue Vec,
170 unsigned IdxVal, SelectionDAG &DAG,
172 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
173 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
176 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
177 /// instructions. This is used because creating CONCAT_VECTOR nodes of
178 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
179 /// large BUILD_VECTORS.
180 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
181 unsigned NumElems, SelectionDAG &DAG,
183 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
184 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
187 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
188 unsigned NumElems, SelectionDAG &DAG,
190 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
191 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
194 static TargetLoweringObjectFile *createTLOF(const Triple &TT) {
195 if (TT.isOSBinFormatMachO()) {
196 if (TT.getArch() == Triple::x86_64)
197 return new X86_64MachoTargetObjectFile();
198 return new TargetLoweringObjectFileMachO();
202 return new X86LinuxTargetObjectFile();
203 if (TT.isOSBinFormatELF())
204 return new TargetLoweringObjectFileELF();
205 if (TT.isKnownWindowsMSVCEnvironment())
206 return new X86WindowsTargetObjectFile();
207 if (TT.isOSBinFormatCOFF())
208 return new TargetLoweringObjectFileCOFF();
209 llvm_unreachable("unknown subtarget type");
212 // FIXME: This should stop caching the target machine as soon as
213 // we can remove resetOperationActions et al.
214 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
215 : TargetLowering(TM, createTLOF(Triple(TM.getTargetTriple()))) {
216 Subtarget = &TM.getSubtarget<X86Subtarget>();
217 X86ScalarSSEf64 = Subtarget->hasSSE2();
218 X86ScalarSSEf32 = Subtarget->hasSSE1();
219 TD = getDataLayout();
221 resetOperationActions();
224 void X86TargetLowering::resetOperationActions() {
225 const TargetMachine &TM = getTargetMachine();
226 static bool FirstTimeThrough = true;
228 // If none of the target options have changed, then we don't need to reset the
229 // operation actions.
230 if (!FirstTimeThrough && TO == TM.Options) return;
232 if (!FirstTimeThrough) {
233 // Reinitialize the actions.
235 FirstTimeThrough = false;
240 // Set up the TargetLowering object.
241 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
243 // X86 is weird, it always uses i8 for shift amounts and setcc results.
244 setBooleanContents(ZeroOrOneBooleanContent);
245 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
246 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
248 // For 64-bit since we have so many registers use the ILP scheduler, for
249 // 32-bit code use the register pressure specific scheduling.
250 // For Atom, always use ILP scheduling.
251 if (Subtarget->isAtom())
252 setSchedulingPreference(Sched::ILP);
253 else if (Subtarget->is64Bit())
254 setSchedulingPreference(Sched::ILP);
256 setSchedulingPreference(Sched::RegPressure);
257 const X86RegisterInfo *RegInfo =
258 static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
259 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
261 // Bypass expensive divides on Atom when compiling with O2
262 if (Subtarget->hasSlowDivide() && TM.getOptLevel() >= CodeGenOpt::Default) {
263 addBypassSlowDiv(32, 8);
264 if (Subtarget->is64Bit())
265 addBypassSlowDiv(64, 16);
268 if (Subtarget->isTargetKnownWindowsMSVC()) {
269 // Setup Windows compiler runtime calls.
270 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
271 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
272 setLibcallName(RTLIB::SREM_I64, "_allrem");
273 setLibcallName(RTLIB::UREM_I64, "_aullrem");
274 setLibcallName(RTLIB::MUL_I64, "_allmul");
275 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
276 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
277 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
278 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
279 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
281 // The _ftol2 runtime function has an unusual calling conv, which
282 // is modeled by a special pseudo-instruction.
283 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
284 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
285 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
286 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
289 if (Subtarget->isTargetDarwin()) {
290 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
291 setUseUnderscoreSetJmp(false);
292 setUseUnderscoreLongJmp(false);
293 } else if (Subtarget->isTargetWindowsGNU()) {
294 // MS runtime is weird: it exports _setjmp, but longjmp!
295 setUseUnderscoreSetJmp(true);
296 setUseUnderscoreLongJmp(false);
298 setUseUnderscoreSetJmp(true);
299 setUseUnderscoreLongJmp(true);
302 // Set up the register classes.
303 addRegisterClass(MVT::i8, &X86::GR8RegClass);
304 addRegisterClass(MVT::i16, &X86::GR16RegClass);
305 addRegisterClass(MVT::i32, &X86::GR32RegClass);
306 if (Subtarget->is64Bit())
307 addRegisterClass(MVT::i64, &X86::GR64RegClass);
309 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
311 // We don't accept any truncstore of integer registers.
312 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
313 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
314 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
315 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
316 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
317 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
319 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
321 // SETOEQ and SETUNE require checking two conditions.
322 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
323 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
324 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
325 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
326 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
327 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
329 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
331 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
332 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
333 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
335 if (Subtarget->is64Bit()) {
336 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
337 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
338 } else if (!TM.Options.UseSoftFloat) {
339 // We have an algorithm for SSE2->double, and we turn this into a
340 // 64-bit FILD followed by conditional FADD for other targets.
341 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
342 // We have an algorithm for SSE2, and we turn this into a 64-bit
343 // FILD for other targets.
344 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
347 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
349 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
350 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
352 if (!TM.Options.UseSoftFloat) {
353 // SSE has no i16 to fp conversion, only i32
354 if (X86ScalarSSEf32) {
355 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
356 // f32 and f64 cases are Legal, f80 case is not
357 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
359 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
360 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
363 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
364 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
367 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
368 // are Legal, f80 is custom lowered.
369 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
370 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
372 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
374 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
375 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
377 if (X86ScalarSSEf32) {
378 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
379 // f32 and f64 cases are Legal, f80 case is not
380 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
382 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
383 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
386 // Handle FP_TO_UINT by promoting the destination to a larger signed
388 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
389 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
390 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
392 if (Subtarget->is64Bit()) {
393 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
394 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
395 } else if (!TM.Options.UseSoftFloat) {
396 // Since AVX is a superset of SSE3, only check for SSE here.
397 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
398 // Expand FP_TO_UINT into a select.
399 // FIXME: We would like to use a Custom expander here eventually to do
400 // the optimal thing for SSE vs. the default expansion in the legalizer.
401 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
403 // With SSE3 we can use fisttpll to convert to a signed i64; without
404 // SSE, we're stuck with a fistpll.
405 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
408 if (isTargetFTOL()) {
409 // Use the _ftol2 runtime function, which has a pseudo-instruction
410 // to handle its weird calling convention.
411 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
414 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
415 if (!X86ScalarSSEf64) {
416 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
417 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
418 if (Subtarget->is64Bit()) {
419 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
420 // Without SSE, i64->f64 goes through memory.
421 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
425 // Scalar integer divide and remainder are lowered to use operations that
426 // produce two results, to match the available instructions. This exposes
427 // the two-result form to trivial CSE, which is able to combine x/y and x%y
428 // into a single instruction.
430 // Scalar integer multiply-high is also lowered to use two-result
431 // operations, to match the available instructions. However, plain multiply
432 // (low) operations are left as Legal, as there are single-result
433 // instructions for this in x86. Using the two-result multiply instructions
434 // when both high and low results are needed must be arranged by dagcombine.
435 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
437 setOperationAction(ISD::MULHS, VT, Expand);
438 setOperationAction(ISD::MULHU, VT, Expand);
439 setOperationAction(ISD::SDIV, VT, Expand);
440 setOperationAction(ISD::UDIV, VT, Expand);
441 setOperationAction(ISD::SREM, VT, Expand);
442 setOperationAction(ISD::UREM, VT, Expand);
444 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
445 setOperationAction(ISD::ADDC, VT, Custom);
446 setOperationAction(ISD::ADDE, VT, Custom);
447 setOperationAction(ISD::SUBC, VT, Custom);
448 setOperationAction(ISD::SUBE, VT, Custom);
451 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
452 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
453 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
454 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
455 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
456 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
457 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
458 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
459 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
460 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
461 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
462 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
463 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
464 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
465 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
466 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
467 if (Subtarget->is64Bit())
468 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
469 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
470 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
471 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
472 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
473 setOperationAction(ISD::FREM , MVT::f32 , Expand);
474 setOperationAction(ISD::FREM , MVT::f64 , Expand);
475 setOperationAction(ISD::FREM , MVT::f80 , Expand);
476 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
478 // Promote the i8 variants and force them on up to i32 which has a shorter
480 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
481 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
482 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
483 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
484 if (Subtarget->hasBMI()) {
485 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
486 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
487 if (Subtarget->is64Bit())
488 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
490 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
491 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
492 if (Subtarget->is64Bit())
493 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
496 if (Subtarget->hasLZCNT()) {
497 // When promoting the i8 variants, force them to i32 for a shorter
499 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
500 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
501 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
502 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
503 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
504 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
505 if (Subtarget->is64Bit())
506 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
508 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
509 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
510 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
511 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
512 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
513 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
514 if (Subtarget->is64Bit()) {
515 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
516 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
520 // Special handling for half-precision floating point conversions.
521 // If we don't have F16C support, then lower half float conversions
522 // into library calls.
523 if (TM.Options.UseSoftFloat || !Subtarget->hasF16C()) {
524 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
525 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
528 // There's never any support for operations beyond MVT::f32.
529 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
530 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
531 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
532 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
534 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
535 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
536 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
537 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
539 if (Subtarget->hasPOPCNT()) {
540 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
542 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
543 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
544 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
545 if (Subtarget->is64Bit())
546 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
549 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
551 if (!Subtarget->hasMOVBE())
552 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
554 // These should be promoted to a larger select which is supported.
555 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
556 // X86 wants to expand cmov itself.
557 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
558 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
559 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
560 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
561 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
562 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
563 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
564 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
565 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
566 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
567 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
568 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
569 if (Subtarget->is64Bit()) {
570 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
571 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
573 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
574 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
575 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
576 // support continuation, user-level threading, and etc.. As a result, no
577 // other SjLj exception interfaces are implemented and please don't build
578 // your own exception handling based on them.
579 // LLVM/Clang supports zero-cost DWARF exception handling.
580 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
581 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
584 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
585 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
586 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
587 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
588 if (Subtarget->is64Bit())
589 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
590 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
591 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
592 if (Subtarget->is64Bit()) {
593 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
594 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
595 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
596 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
597 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
599 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
600 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
601 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
602 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
603 if (Subtarget->is64Bit()) {
604 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
605 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
606 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
609 if (Subtarget->hasSSE1())
610 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
612 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
614 // Expand certain atomics
615 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
617 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
618 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
619 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
622 if (Subtarget->hasCmpxchg16b()) {
623 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
626 // FIXME - use subtarget debug flags
627 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
628 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
629 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
632 if (Subtarget->is64Bit()) {
633 setExceptionPointerRegister(X86::RAX);
634 setExceptionSelectorRegister(X86::RDX);
636 setExceptionPointerRegister(X86::EAX);
637 setExceptionSelectorRegister(X86::EDX);
639 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
640 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
642 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
643 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
645 setOperationAction(ISD::TRAP, MVT::Other, Legal);
646 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
648 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
649 setOperationAction(ISD::VASTART , MVT::Other, Custom);
650 setOperationAction(ISD::VAEND , MVT::Other, Expand);
651 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
652 // TargetInfo::X86_64ABIBuiltinVaList
653 setOperationAction(ISD::VAARG , MVT::Other, Custom);
654 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
656 // TargetInfo::CharPtrBuiltinVaList
657 setOperationAction(ISD::VAARG , MVT::Other, Expand);
658 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
661 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
662 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
664 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
665 MVT::i64 : MVT::i32, Custom);
667 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
668 // f32 and f64 use SSE.
669 // Set up the FP register classes.
670 addRegisterClass(MVT::f32, &X86::FR32RegClass);
671 addRegisterClass(MVT::f64, &X86::FR64RegClass);
673 // Use ANDPD to simulate FABS.
674 setOperationAction(ISD::FABS , MVT::f64, Custom);
675 setOperationAction(ISD::FABS , MVT::f32, Custom);
677 // Use XORP to simulate FNEG.
678 setOperationAction(ISD::FNEG , MVT::f64, Custom);
679 setOperationAction(ISD::FNEG , MVT::f32, Custom);
681 // Use ANDPD and ORPD to simulate FCOPYSIGN.
682 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
683 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
685 // Lower this to FGETSIGNx86 plus an AND.
686 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
687 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
689 // We don't support sin/cos/fmod
690 setOperationAction(ISD::FSIN , MVT::f64, Expand);
691 setOperationAction(ISD::FCOS , MVT::f64, Expand);
692 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
693 setOperationAction(ISD::FSIN , MVT::f32, Expand);
694 setOperationAction(ISD::FCOS , MVT::f32, Expand);
695 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
697 // Expand FP immediates into loads from the stack, except for the special
699 addLegalFPImmediate(APFloat(+0.0)); // xorpd
700 addLegalFPImmediate(APFloat(+0.0f)); // xorps
701 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
702 // Use SSE for f32, x87 for f64.
703 // Set up the FP register classes.
704 addRegisterClass(MVT::f32, &X86::FR32RegClass);
705 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
707 // Use ANDPS to simulate FABS.
708 setOperationAction(ISD::FABS , MVT::f32, Custom);
710 // Use XORP to simulate FNEG.
711 setOperationAction(ISD::FNEG , MVT::f32, Custom);
713 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
715 // Use ANDPS and ORPS to simulate FCOPYSIGN.
716 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
717 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
719 // We don't support sin/cos/fmod
720 setOperationAction(ISD::FSIN , MVT::f32, Expand);
721 setOperationAction(ISD::FCOS , MVT::f32, Expand);
722 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
724 // Special cases we handle for FP constants.
725 addLegalFPImmediate(APFloat(+0.0f)); // xorps
726 addLegalFPImmediate(APFloat(+0.0)); // FLD0
727 addLegalFPImmediate(APFloat(+1.0)); // FLD1
728 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
729 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
731 if (!TM.Options.UnsafeFPMath) {
732 setOperationAction(ISD::FSIN , MVT::f64, Expand);
733 setOperationAction(ISD::FCOS , MVT::f64, Expand);
734 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
736 } else if (!TM.Options.UseSoftFloat) {
737 // f32 and f64 in x87.
738 // Set up the FP register classes.
739 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
740 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
742 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
743 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
744 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
745 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
747 if (!TM.Options.UnsafeFPMath) {
748 setOperationAction(ISD::FSIN , MVT::f64, Expand);
749 setOperationAction(ISD::FSIN , MVT::f32, Expand);
750 setOperationAction(ISD::FCOS , MVT::f64, Expand);
751 setOperationAction(ISD::FCOS , MVT::f32, Expand);
752 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
753 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
755 addLegalFPImmediate(APFloat(+0.0)); // FLD0
756 addLegalFPImmediate(APFloat(+1.0)); // FLD1
757 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
758 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
759 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
760 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
761 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
762 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
765 // We don't support FMA.
766 setOperationAction(ISD::FMA, MVT::f64, Expand);
767 setOperationAction(ISD::FMA, MVT::f32, Expand);
769 // Long double always uses X87.
770 if (!TM.Options.UseSoftFloat) {
771 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
772 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
773 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
775 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
776 addLegalFPImmediate(TmpFlt); // FLD0
778 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
781 APFloat TmpFlt2(+1.0);
782 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
784 addLegalFPImmediate(TmpFlt2); // FLD1
785 TmpFlt2.changeSign();
786 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
789 if (!TM.Options.UnsafeFPMath) {
790 setOperationAction(ISD::FSIN , MVT::f80, Expand);
791 setOperationAction(ISD::FCOS , MVT::f80, Expand);
792 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
795 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
796 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
797 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
798 setOperationAction(ISD::FRINT, MVT::f80, Expand);
799 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
800 setOperationAction(ISD::FMA, MVT::f80, Expand);
803 // Always use a library call for pow.
804 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
805 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
806 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
808 setOperationAction(ISD::FLOG, MVT::f80, Expand);
809 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
810 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
811 setOperationAction(ISD::FEXP, MVT::f80, Expand);
812 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
814 // First set operation action for all vector types to either promote
815 // (for widening) or expand (for scalarization). Then we will selectively
816 // turn on ones that can be effectively codegen'd.
817 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
818 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
819 MVT VT = (MVT::SimpleValueType)i;
820 setOperationAction(ISD::ADD , VT, Expand);
821 setOperationAction(ISD::SUB , VT, Expand);
822 setOperationAction(ISD::FADD, VT, Expand);
823 setOperationAction(ISD::FNEG, VT, Expand);
824 setOperationAction(ISD::FSUB, VT, Expand);
825 setOperationAction(ISD::MUL , VT, Expand);
826 setOperationAction(ISD::FMUL, VT, Expand);
827 setOperationAction(ISD::SDIV, VT, Expand);
828 setOperationAction(ISD::UDIV, VT, Expand);
829 setOperationAction(ISD::FDIV, VT, Expand);
830 setOperationAction(ISD::SREM, VT, Expand);
831 setOperationAction(ISD::UREM, VT, Expand);
832 setOperationAction(ISD::LOAD, VT, Expand);
833 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
834 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
835 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
836 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
837 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
838 setOperationAction(ISD::FABS, VT, Expand);
839 setOperationAction(ISD::FSIN, VT, Expand);
840 setOperationAction(ISD::FSINCOS, VT, Expand);
841 setOperationAction(ISD::FCOS, VT, Expand);
842 setOperationAction(ISD::FSINCOS, VT, Expand);
843 setOperationAction(ISD::FREM, VT, Expand);
844 setOperationAction(ISD::FMA, VT, Expand);
845 setOperationAction(ISD::FPOWI, VT, Expand);
846 setOperationAction(ISD::FSQRT, VT, Expand);
847 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
848 setOperationAction(ISD::FFLOOR, VT, Expand);
849 setOperationAction(ISD::FCEIL, VT, Expand);
850 setOperationAction(ISD::FTRUNC, VT, Expand);
851 setOperationAction(ISD::FRINT, VT, Expand);
852 setOperationAction(ISD::FNEARBYINT, VT, Expand);
853 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
854 setOperationAction(ISD::MULHS, VT, Expand);
855 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
856 setOperationAction(ISD::MULHU, VT, Expand);
857 setOperationAction(ISD::SDIVREM, VT, Expand);
858 setOperationAction(ISD::UDIVREM, VT, Expand);
859 setOperationAction(ISD::FPOW, VT, Expand);
860 setOperationAction(ISD::CTPOP, VT, Expand);
861 setOperationAction(ISD::CTTZ, VT, Expand);
862 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
863 setOperationAction(ISD::CTLZ, VT, Expand);
864 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
865 setOperationAction(ISD::SHL, VT, Expand);
866 setOperationAction(ISD::SRA, VT, Expand);
867 setOperationAction(ISD::SRL, VT, Expand);
868 setOperationAction(ISD::ROTL, VT, Expand);
869 setOperationAction(ISD::ROTR, VT, Expand);
870 setOperationAction(ISD::BSWAP, VT, Expand);
871 setOperationAction(ISD::SETCC, VT, Expand);
872 setOperationAction(ISD::FLOG, VT, Expand);
873 setOperationAction(ISD::FLOG2, VT, Expand);
874 setOperationAction(ISD::FLOG10, VT, Expand);
875 setOperationAction(ISD::FEXP, VT, Expand);
876 setOperationAction(ISD::FEXP2, VT, Expand);
877 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
878 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
879 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
880 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
881 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
882 setOperationAction(ISD::TRUNCATE, VT, Expand);
883 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
884 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
885 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
886 setOperationAction(ISD::VSELECT, VT, Expand);
887 setOperationAction(ISD::SELECT_CC, VT, Expand);
888 for (int InnerVT = MVT::FIRST_VECTOR_VALUETYPE;
889 InnerVT <= MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
890 setTruncStoreAction(VT,
891 (MVT::SimpleValueType)InnerVT, Expand);
892 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
893 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
895 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like types,
896 // we have to deal with them whether we ask for Expansion or not. Setting
897 // Expand causes its own optimisation problems though, so leave them legal.
898 if (VT.getVectorElementType() == MVT::i1)
899 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
902 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
903 // with -msoft-float, disable use of MMX as well.
904 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
905 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
906 // No operations on x86mmx supported, everything uses intrinsics.
909 // MMX-sized vectors (other than x86mmx) are expected to be expanded
910 // into smaller operations.
911 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
912 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
913 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
914 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
915 setOperationAction(ISD::AND, MVT::v8i8, Expand);
916 setOperationAction(ISD::AND, MVT::v4i16, Expand);
917 setOperationAction(ISD::AND, MVT::v2i32, Expand);
918 setOperationAction(ISD::AND, MVT::v1i64, Expand);
919 setOperationAction(ISD::OR, MVT::v8i8, Expand);
920 setOperationAction(ISD::OR, MVT::v4i16, Expand);
921 setOperationAction(ISD::OR, MVT::v2i32, Expand);
922 setOperationAction(ISD::OR, MVT::v1i64, Expand);
923 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
924 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
925 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
926 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
927 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
928 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
929 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
930 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
931 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
932 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
933 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
934 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
935 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
936 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
937 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
938 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
939 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
941 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
942 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
944 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
945 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
946 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
947 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
948 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
949 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
950 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
951 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
952 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
953 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
954 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
955 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
958 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
959 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
961 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
962 // registers cannot be used even for integer operations.
963 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
964 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
965 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
966 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
968 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
969 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
970 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
971 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
972 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
973 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
974 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
975 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
976 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
977 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
978 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
979 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
980 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
981 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
982 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
983 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
984 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
985 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
986 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
987 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
988 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
989 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
991 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
992 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
993 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
994 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
996 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
997 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
998 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
999 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1000 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1002 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
1003 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1004 MVT VT = (MVT::SimpleValueType)i;
1005 // Do not attempt to custom lower non-power-of-2 vectors
1006 if (!isPowerOf2_32(VT.getVectorNumElements()))
1008 // Do not attempt to custom lower non-128-bit vectors
1009 if (!VT.is128BitVector())
1011 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1012 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1013 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1016 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
1017 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
1018 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
1019 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
1020 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
1021 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
1023 if (Subtarget->is64Bit()) {
1024 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1025 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1028 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
1029 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1030 MVT VT = (MVT::SimpleValueType)i;
1032 // Do not attempt to promote non-128-bit vectors
1033 if (!VT.is128BitVector())
1036 setOperationAction(ISD::AND, VT, Promote);
1037 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
1038 setOperationAction(ISD::OR, VT, Promote);
1039 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
1040 setOperationAction(ISD::XOR, VT, Promote);
1041 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
1042 setOperationAction(ISD::LOAD, VT, Promote);
1043 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
1044 setOperationAction(ISD::SELECT, VT, Promote);
1045 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
1048 // Custom lower v2i64 and v2f64 selects.
1049 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
1050 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
1051 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
1052 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
1054 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1055 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1057 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
1058 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
1059 // As there is no 64-bit GPR available, we need build a special custom
1060 // sequence to convert from v2i32 to v2f32.
1061 if (!Subtarget->is64Bit())
1062 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
1064 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
1065 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
1067 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, Legal);
1069 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
1070 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
1071 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
1074 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) {
1075 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
1076 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
1077 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
1078 setOperationAction(ISD::FRINT, MVT::f32, Legal);
1079 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1080 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
1081 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1082 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1083 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1084 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1086 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
1087 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
1088 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
1089 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1090 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1091 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1092 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1093 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1094 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1095 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1097 // FIXME: Do we need to handle scalar-to-vector here?
1098 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
1100 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
1101 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
1102 setOperationAction(ISD::VSELECT, MVT::v4i32, Custom);
1103 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
1104 setOperationAction(ISD::VSELECT, MVT::v8i16, Custom);
1105 // There is no BLENDI for byte vectors. We don't need to custom lower
1106 // some vselects for now.
1107 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
1109 // i8 and i16 vectors are custom , because the source register and source
1110 // source memory operand types are not the same width. f32 vectors are
1111 // custom since the immediate controlling the insert encodes additional
1113 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1114 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1115 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1116 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1118 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1119 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1120 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1121 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1123 // FIXME: these should be Legal but thats only for the case where
1124 // the index is constant. For now custom expand to deal with that.
1125 if (Subtarget->is64Bit()) {
1126 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1127 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1131 if (Subtarget->hasSSE2()) {
1132 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1133 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1135 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1136 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1138 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1139 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1141 // In the customized shift lowering, the legal cases in AVX2 will be
1143 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1144 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1146 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1147 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1149 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1152 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1153 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1154 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1155 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1156 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1157 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1158 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1160 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1161 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1162 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1164 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1165 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1166 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1167 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1168 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1169 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1170 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1171 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1172 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1173 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1174 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1175 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1177 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1178 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1179 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1180 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1181 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1182 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1183 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1184 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1185 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1186 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1187 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1188 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1190 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1191 // even though v8i16 is a legal type.
1192 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1193 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1194 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1196 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1197 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1198 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1200 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1201 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1203 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, Legal);
1205 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1206 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1208 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1209 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1211 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1212 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1214 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1215 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1216 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1217 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1219 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1220 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1221 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1223 setOperationAction(ISD::VSELECT, MVT::v4f64, Custom);
1224 setOperationAction(ISD::VSELECT, MVT::v4i64, Custom);
1225 setOperationAction(ISD::VSELECT, MVT::v8i32, Custom);
1226 setOperationAction(ISD::VSELECT, MVT::v8f32, Custom);
1228 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1229 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1230 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1231 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1232 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1233 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1234 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1235 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1236 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1237 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1238 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1239 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1241 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1242 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1243 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1244 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1245 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1246 setOperationAction(ISD::FMA, MVT::f32, Legal);
1247 setOperationAction(ISD::FMA, MVT::f64, Legal);
1250 if (Subtarget->hasInt256()) {
1251 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1252 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1253 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1254 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1256 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1257 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1258 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1259 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1261 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1262 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1263 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1264 // Don't lower v32i8 because there is no 128-bit byte mul
1266 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1267 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1268 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1269 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1271 setOperationAction(ISD::VSELECT, MVT::v16i16, Custom);
1272 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1274 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1275 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1276 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1277 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1279 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1280 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1281 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1282 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1284 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1285 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1286 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1287 // Don't lower v32i8 because there is no 128-bit byte mul
1290 // In the customized shift lowering, the legal cases in AVX2 will be
1292 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1293 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1295 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1296 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1298 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1300 // Custom lower several nodes for 256-bit types.
1301 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
1302 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
1303 MVT VT = (MVT::SimpleValueType)i;
1305 // Extract subvector is special because the value type
1306 // (result) is 128-bit but the source is 256-bit wide.
1307 if (VT.is128BitVector())
1308 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1310 // Do not attempt to custom lower other non-256-bit vectors
1311 if (!VT.is256BitVector())
1314 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1315 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1316 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1317 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1318 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1319 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1320 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1323 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1324 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1325 MVT VT = (MVT::SimpleValueType)i;
1327 // Do not attempt to promote non-256-bit vectors
1328 if (!VT.is256BitVector())
1331 setOperationAction(ISD::AND, VT, Promote);
1332 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1333 setOperationAction(ISD::OR, VT, Promote);
1334 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1335 setOperationAction(ISD::XOR, VT, Promote);
1336 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1337 setOperationAction(ISD::LOAD, VT, Promote);
1338 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1339 setOperationAction(ISD::SELECT, VT, Promote);
1340 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1344 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) {
1345 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1346 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1347 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1348 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1350 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1351 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1352 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1354 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1355 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1356 setOperationAction(ISD::XOR, MVT::i1, Legal);
1357 setOperationAction(ISD::OR, MVT::i1, Legal);
1358 setOperationAction(ISD::AND, MVT::i1, Legal);
1359 setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, Legal);
1360 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1361 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1362 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1363 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1364 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1366 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1367 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1368 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1369 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1370 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1371 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1373 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1374 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1375 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1376 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1377 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1378 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1379 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1380 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1382 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1383 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1384 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1385 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1386 if (Subtarget->is64Bit()) {
1387 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1388 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1389 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1390 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1392 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1393 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1394 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1395 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1396 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1397 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1398 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1399 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1400 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1401 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1403 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1404 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1405 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1406 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1407 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1408 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1409 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1410 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1411 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1412 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1413 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1414 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1415 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1417 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1418 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1419 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1420 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1421 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1422 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1424 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1425 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1427 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1429 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1430 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1431 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1432 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1433 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1434 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1435 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1436 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1437 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1439 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1440 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1442 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1443 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1445 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1447 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1448 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1450 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1451 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1453 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1454 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1456 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1457 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1458 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1459 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1460 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1461 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1463 if (Subtarget->hasCDI()) {
1464 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1465 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1468 // Custom lower several nodes.
1469 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
1470 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
1471 MVT VT = (MVT::SimpleValueType)i;
1473 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1474 // Extract subvector is special because the value type
1475 // (result) is 256/128-bit but the source is 512-bit wide.
1476 if (VT.is128BitVector() || VT.is256BitVector())
1477 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1479 if (VT.getVectorElementType() == MVT::i1)
1480 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1482 // Do not attempt to custom lower other non-512-bit vectors
1483 if (!VT.is512BitVector())
1486 if ( EltSize >= 32) {
1487 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1488 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1489 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1490 setOperationAction(ISD::VSELECT, VT, Legal);
1491 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1492 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1493 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1496 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1497 MVT VT = (MVT::SimpleValueType)i;
1499 // Do not attempt to promote non-256-bit vectors
1500 if (!VT.is512BitVector())
1503 setOperationAction(ISD::SELECT, VT, Promote);
1504 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1508 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1509 // of this type with custom code.
1510 for (int VT = MVT::FIRST_VECTOR_VALUETYPE;
1511 VT != MVT::LAST_VECTOR_VALUETYPE; VT++) {
1512 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
1516 // We want to custom lower some of our intrinsics.
1517 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1518 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1519 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1520 if (!Subtarget->is64Bit())
1521 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1523 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1524 // handle type legalization for these operations here.
1526 // FIXME: We really should do custom legalization for addition and
1527 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1528 // than generic legalization for 64-bit multiplication-with-overflow, though.
1529 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1530 // Add/Sub/Mul with overflow operations are custom lowered.
1532 setOperationAction(ISD::SADDO, VT, Custom);
1533 setOperationAction(ISD::UADDO, VT, Custom);
1534 setOperationAction(ISD::SSUBO, VT, Custom);
1535 setOperationAction(ISD::USUBO, VT, Custom);
1536 setOperationAction(ISD::SMULO, VT, Custom);
1537 setOperationAction(ISD::UMULO, VT, Custom);
1540 // There are no 8-bit 3-address imul/mul instructions
1541 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1542 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1544 if (!Subtarget->is64Bit()) {
1545 // These libcalls are not available in 32-bit.
1546 setLibcallName(RTLIB::SHL_I128, nullptr);
1547 setLibcallName(RTLIB::SRL_I128, nullptr);
1548 setLibcallName(RTLIB::SRA_I128, nullptr);
1551 // Combine sin / cos into one node or libcall if possible.
1552 if (Subtarget->hasSinCos()) {
1553 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1554 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1555 if (Subtarget->isTargetDarwin()) {
1556 // For MacOSX, we don't want to the normal expansion of a libcall to
1557 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
1559 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1560 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1564 if (Subtarget->isTargetWin64()) {
1565 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1566 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1567 setOperationAction(ISD::SREM, MVT::i128, Custom);
1568 setOperationAction(ISD::UREM, MVT::i128, Custom);
1569 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1570 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1573 // We have target-specific dag combine patterns for the following nodes:
1574 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1575 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1576 setTargetDAGCombine(ISD::VSELECT);
1577 setTargetDAGCombine(ISD::SELECT);
1578 setTargetDAGCombine(ISD::SHL);
1579 setTargetDAGCombine(ISD::SRA);
1580 setTargetDAGCombine(ISD::SRL);
1581 setTargetDAGCombine(ISD::OR);
1582 setTargetDAGCombine(ISD::AND);
1583 setTargetDAGCombine(ISD::ADD);
1584 setTargetDAGCombine(ISD::FADD);
1585 setTargetDAGCombine(ISD::FSUB);
1586 setTargetDAGCombine(ISD::FMA);
1587 setTargetDAGCombine(ISD::SUB);
1588 setTargetDAGCombine(ISD::LOAD);
1589 setTargetDAGCombine(ISD::STORE);
1590 setTargetDAGCombine(ISD::ZERO_EXTEND);
1591 setTargetDAGCombine(ISD::ANY_EXTEND);
1592 setTargetDAGCombine(ISD::SIGN_EXTEND);
1593 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1594 setTargetDAGCombine(ISD::TRUNCATE);
1595 setTargetDAGCombine(ISD::SINT_TO_FP);
1596 setTargetDAGCombine(ISD::SETCC);
1597 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1598 setTargetDAGCombine(ISD::BUILD_VECTOR);
1599 if (Subtarget->is64Bit())
1600 setTargetDAGCombine(ISD::MUL);
1601 setTargetDAGCombine(ISD::XOR);
1603 computeRegisterProperties();
1605 // On Darwin, -Os means optimize for size without hurting performance,
1606 // do not reduce the limit.
1607 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1608 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1609 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1610 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1611 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1612 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1613 setPrefLoopAlignment(4); // 2^4 bytes.
1615 // Predictable cmov don't hurt on atom because it's in-order.
1616 PredictableSelectIsExpensive = !Subtarget->isAtom();
1618 setPrefFunctionAlignment(4); // 2^4 bytes.
1621 TargetLoweringBase::LegalizeTypeAction
1622 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1623 if (ExperimentalVectorWideningLegalization &&
1624 VT.getVectorNumElements() != 1 &&
1625 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1626 return TypeWidenVector;
1628 return TargetLoweringBase::getPreferredVectorAction(VT);
1631 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1633 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1635 if (Subtarget->hasAVX512())
1636 switch(VT.getVectorNumElements()) {
1637 case 8: return MVT::v8i1;
1638 case 16: return MVT::v16i1;
1641 return VT.changeVectorElementTypeToInteger();
1644 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1645 /// the desired ByVal argument alignment.
1646 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1649 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1650 if (VTy->getBitWidth() == 128)
1652 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1653 unsigned EltAlign = 0;
1654 getMaxByValAlign(ATy->getElementType(), EltAlign);
1655 if (EltAlign > MaxAlign)
1656 MaxAlign = EltAlign;
1657 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1658 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1659 unsigned EltAlign = 0;
1660 getMaxByValAlign(STy->getElementType(i), EltAlign);
1661 if (EltAlign > MaxAlign)
1662 MaxAlign = EltAlign;
1669 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1670 /// function arguments in the caller parameter area. For X86, aggregates
1671 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1672 /// are at 4-byte boundaries.
1673 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1674 if (Subtarget->is64Bit()) {
1675 // Max of 8 and alignment of type.
1676 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1683 if (Subtarget->hasSSE1())
1684 getMaxByValAlign(Ty, Align);
1688 /// getOptimalMemOpType - Returns the target specific optimal type for load
1689 /// and store operations as a result of memset, memcpy, and memmove
1690 /// lowering. If DstAlign is zero that means it's safe to destination
1691 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1692 /// means there isn't a need to check it against alignment requirement,
1693 /// probably because the source does not need to be loaded. If 'IsMemset' is
1694 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1695 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1696 /// source is constant so it does not need to be loaded.
1697 /// It returns EVT::Other if the type should be determined using generic
1698 /// target-independent logic.
1700 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1701 unsigned DstAlign, unsigned SrcAlign,
1702 bool IsMemset, bool ZeroMemset,
1704 MachineFunction &MF) const {
1705 const Function *F = MF.getFunction();
1706 if ((!IsMemset || ZeroMemset) &&
1707 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1708 Attribute::NoImplicitFloat)) {
1710 (Subtarget->isUnalignedMemAccessFast() ||
1711 ((DstAlign == 0 || DstAlign >= 16) &&
1712 (SrcAlign == 0 || SrcAlign >= 16)))) {
1714 if (Subtarget->hasInt256())
1716 if (Subtarget->hasFp256())
1719 if (Subtarget->hasSSE2())
1721 if (Subtarget->hasSSE1())
1723 } else if (!MemcpyStrSrc && Size >= 8 &&
1724 !Subtarget->is64Bit() &&
1725 Subtarget->hasSSE2()) {
1726 // Do not use f64 to lower memcpy if source is string constant. It's
1727 // better to use i32 to avoid the loads.
1731 if (Subtarget->is64Bit() && Size >= 8)
1736 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1738 return X86ScalarSSEf32;
1739 else if (VT == MVT::f64)
1740 return X86ScalarSSEf64;
1745 X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
1749 *Fast = Subtarget->isUnalignedMemAccessFast();
1753 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1754 /// current function. The returned value is a member of the
1755 /// MachineJumpTableInfo::JTEntryKind enum.
1756 unsigned X86TargetLowering::getJumpTableEncoding() const {
1757 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1759 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1760 Subtarget->isPICStyleGOT())
1761 return MachineJumpTableInfo::EK_Custom32;
1763 // Otherwise, use the normal jump table encoding heuristics.
1764 return TargetLowering::getJumpTableEncoding();
1768 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1769 const MachineBasicBlock *MBB,
1770 unsigned uid,MCContext &Ctx) const{
1771 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1772 Subtarget->isPICStyleGOT());
1773 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1775 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1776 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1779 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1781 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1782 SelectionDAG &DAG) const {
1783 if (!Subtarget->is64Bit())
1784 // This doesn't have SDLoc associated with it, but is not really the
1785 // same as a Register.
1786 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1790 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1791 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1793 const MCExpr *X86TargetLowering::
1794 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1795 MCContext &Ctx) const {
1796 // X86-64 uses RIP relative addressing based on the jump table label.
1797 if (Subtarget->isPICStyleRIPRel())
1798 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1800 // Otherwise, the reference is relative to the PIC base.
1801 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1804 // FIXME: Why this routine is here? Move to RegInfo!
1805 std::pair<const TargetRegisterClass*, uint8_t>
1806 X86TargetLowering::findRepresentativeClass(MVT VT) const{
1807 const TargetRegisterClass *RRC = nullptr;
1809 switch (VT.SimpleTy) {
1811 return TargetLowering::findRepresentativeClass(VT);
1812 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1813 RRC = Subtarget->is64Bit() ?
1814 (const TargetRegisterClass*)&X86::GR64RegClass :
1815 (const TargetRegisterClass*)&X86::GR32RegClass;
1818 RRC = &X86::VR64RegClass;
1820 case MVT::f32: case MVT::f64:
1821 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1822 case MVT::v4f32: case MVT::v2f64:
1823 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1825 RRC = &X86::VR128RegClass;
1828 return std::make_pair(RRC, Cost);
1831 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1832 unsigned &Offset) const {
1833 if (!Subtarget->isTargetLinux())
1836 if (Subtarget->is64Bit()) {
1837 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1839 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1851 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1852 unsigned DestAS) const {
1853 assert(SrcAS != DestAS && "Expected different address spaces!");
1855 return SrcAS < 256 && DestAS < 256;
1858 //===----------------------------------------------------------------------===//
1859 // Return Value Calling Convention Implementation
1860 //===----------------------------------------------------------------------===//
1862 #include "X86GenCallingConv.inc"
1865 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1866 MachineFunction &MF, bool isVarArg,
1867 const SmallVectorImpl<ISD::OutputArg> &Outs,
1868 LLVMContext &Context) const {
1869 SmallVector<CCValAssign, 16> RVLocs;
1870 CCState CCInfo(CallConv, isVarArg, MF, MF.getTarget(),
1872 return CCInfo.CheckReturn(Outs, RetCC_X86);
1875 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
1876 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
1881 X86TargetLowering::LowerReturn(SDValue Chain,
1882 CallingConv::ID CallConv, bool isVarArg,
1883 const SmallVectorImpl<ISD::OutputArg> &Outs,
1884 const SmallVectorImpl<SDValue> &OutVals,
1885 SDLoc dl, SelectionDAG &DAG) const {
1886 MachineFunction &MF = DAG.getMachineFunction();
1887 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1889 SmallVector<CCValAssign, 16> RVLocs;
1890 CCState CCInfo(CallConv, isVarArg, MF, DAG.getTarget(),
1891 RVLocs, *DAG.getContext());
1892 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1895 SmallVector<SDValue, 6> RetOps;
1896 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1897 // Operand #1 = Bytes To Pop
1898 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1901 // Copy the result values into the output registers.
1902 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1903 CCValAssign &VA = RVLocs[i];
1904 assert(VA.isRegLoc() && "Can only return in registers!");
1905 SDValue ValToCopy = OutVals[i];
1906 EVT ValVT = ValToCopy.getValueType();
1908 // Promote values to the appropriate types
1909 if (VA.getLocInfo() == CCValAssign::SExt)
1910 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1911 else if (VA.getLocInfo() == CCValAssign::ZExt)
1912 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1913 else if (VA.getLocInfo() == CCValAssign::AExt)
1914 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1915 else if (VA.getLocInfo() == CCValAssign::BCvt)
1916 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
1918 assert(VA.getLocInfo() != CCValAssign::FPExt &&
1919 "Unexpected FP-extend for return value.");
1921 // If this is x86-64, and we disabled SSE, we can't return FP values,
1922 // or SSE or MMX vectors.
1923 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1924 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1925 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1926 report_fatal_error("SSE register return with SSE disabled");
1928 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1929 // llvm-gcc has never done it right and no one has noticed, so this
1930 // should be OK for now.
1931 if (ValVT == MVT::f64 &&
1932 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1933 report_fatal_error("SSE2 register return with SSE2 disabled");
1935 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1936 // the RET instruction and handled by the FP Stackifier.
1937 if (VA.getLocReg() == X86::ST0 ||
1938 VA.getLocReg() == X86::ST1) {
1939 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1940 // change the value to the FP stack register class.
1941 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1942 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1943 RetOps.push_back(ValToCopy);
1944 // Don't emit a copytoreg.
1948 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1949 // which is returned in RAX / RDX.
1950 if (Subtarget->is64Bit()) {
1951 if (ValVT == MVT::x86mmx) {
1952 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1953 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1954 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1956 // If we don't have SSE2 available, convert to v4f32 so the generated
1957 // register is legal.
1958 if (!Subtarget->hasSSE2())
1959 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1964 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1965 Flag = Chain.getValue(1);
1966 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1969 // The x86-64 ABIs require that for returning structs by value we copy
1970 // the sret argument into %rax/%eax (depending on ABI) for the return.
1971 // Win32 requires us to put the sret argument to %eax as well.
1972 // We saved the argument into a virtual register in the entry block,
1973 // so now we copy the value out and into %rax/%eax.
1974 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
1975 (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) {
1976 MachineFunction &MF = DAG.getMachineFunction();
1977 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1978 unsigned Reg = FuncInfo->getSRetReturnReg();
1980 "SRetReturnReg should have been set in LowerFormalArguments().");
1981 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1984 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
1985 X86::RAX : X86::EAX;
1986 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
1987 Flag = Chain.getValue(1);
1989 // RAX/EAX now acts like a return value.
1990 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
1993 RetOps[0] = Chain; // Update chain.
1995 // Add the flag if we have it.
1997 RetOps.push_back(Flag);
1999 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2002 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2003 if (N->getNumValues() != 1)
2005 if (!N->hasNUsesOfValue(1, 0))
2008 SDValue TCChain = Chain;
2009 SDNode *Copy = *N->use_begin();
2010 if (Copy->getOpcode() == ISD::CopyToReg) {
2011 // If the copy has a glue operand, we conservatively assume it isn't safe to
2012 // perform a tail call.
2013 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2015 TCChain = Copy->getOperand(0);
2016 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2019 bool HasRet = false;
2020 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2022 if (UI->getOpcode() != X86ISD::RET_FLAG)
2035 X86TargetLowering::getTypeForExtArgOrReturn(MVT VT,
2036 ISD::NodeType ExtendKind) const {
2038 // TODO: Is this also valid on 32-bit?
2039 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2040 ReturnMVT = MVT::i8;
2042 ReturnMVT = MVT::i32;
2044 MVT MinVT = getRegisterType(ReturnMVT);
2045 return VT.bitsLT(MinVT) ? MinVT : VT;
2048 /// LowerCallResult - Lower the result values of a call into the
2049 /// appropriate copies out of appropriate physical registers.
2052 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2053 CallingConv::ID CallConv, bool isVarArg,
2054 const SmallVectorImpl<ISD::InputArg> &Ins,
2055 SDLoc dl, SelectionDAG &DAG,
2056 SmallVectorImpl<SDValue> &InVals) const {
2058 // Assign locations to each value returned by this call.
2059 SmallVector<CCValAssign, 16> RVLocs;
2060 bool Is64Bit = Subtarget->is64Bit();
2061 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2062 DAG.getTarget(), RVLocs, *DAG.getContext());
2063 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2065 // Copy all of the result registers out of their specified physreg.
2066 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2067 CCValAssign &VA = RVLocs[i];
2068 EVT CopyVT = VA.getValVT();
2070 // If this is x86-64, and we disabled SSE, we can't return FP values
2071 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2072 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2073 report_fatal_error("SSE register return with SSE disabled");
2078 // If this is a call to a function that returns an fp value on the floating
2079 // point stack, we must guarantee the value is popped from the stack, so
2080 // a CopyFromReg is not good enough - the copy instruction may be eliminated
2081 // if the return value is not used. We use the FpPOP_RETVAL instruction
2083 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
2084 // If we prefer to use the value in xmm registers, copy it out as f80 and
2085 // use a truncate to move it from fp stack reg to xmm reg.
2086 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
2087 SDValue Ops[] = { Chain, InFlag };
2088 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
2089 MVT::Other, MVT::Glue, Ops), 1);
2090 Val = Chain.getValue(0);
2092 // Round the f80 to the right size, which also moves it to the appropriate
2094 if (CopyVT != VA.getValVT())
2095 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2096 // This truncation won't change the value.
2097 DAG.getIntPtrConstant(1));
2099 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2100 CopyVT, InFlag).getValue(1);
2101 Val = Chain.getValue(0);
2103 InFlag = Chain.getValue(2);
2104 InVals.push_back(Val);
2110 //===----------------------------------------------------------------------===//
2111 // C & StdCall & Fast Calling Convention implementation
2112 //===----------------------------------------------------------------------===//
2113 // StdCall calling convention seems to be standard for many Windows' API
2114 // routines and around. It differs from C calling convention just a little:
2115 // callee should clean up the stack, not caller. Symbols should be also
2116 // decorated in some fancy way :) It doesn't support any vector arguments.
2117 // For info on fast calling convention see Fast Calling Convention (tail call)
2118 // implementation LowerX86_32FastCCCallTo.
2120 /// CallIsStructReturn - Determines whether a call uses struct return
2122 enum StructReturnType {
2127 static StructReturnType
2128 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2130 return NotStructReturn;
2132 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2133 if (!Flags.isSRet())
2134 return NotStructReturn;
2135 if (Flags.isInReg())
2136 return RegStructReturn;
2137 return StackStructReturn;
2140 /// ArgsAreStructReturn - Determines whether a function uses struct
2141 /// return semantics.
2142 static StructReturnType
2143 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2145 return NotStructReturn;
2147 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2148 if (!Flags.isSRet())
2149 return NotStructReturn;
2150 if (Flags.isInReg())
2151 return RegStructReturn;
2152 return StackStructReturn;
2155 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2156 /// by "Src" to address "Dst" with size and alignment information specified by
2157 /// the specific parameter attribute. The copy will be passed as a byval
2158 /// function parameter.
2160 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2161 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2163 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2165 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2166 /*isVolatile*/false, /*AlwaysInline=*/true,
2167 MachinePointerInfo(), MachinePointerInfo());
2170 /// IsTailCallConvention - Return true if the calling convention is one that
2171 /// supports tail call optimization.
2172 static bool IsTailCallConvention(CallingConv::ID CC) {
2173 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2174 CC == CallingConv::HiPE);
2177 /// \brief Return true if the calling convention is a C calling convention.
2178 static bool IsCCallConvention(CallingConv::ID CC) {
2179 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2180 CC == CallingConv::X86_64_SysV);
2183 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2184 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2188 CallingConv::ID CalleeCC = CS.getCallingConv();
2189 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2195 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
2196 /// a tailcall target by changing its ABI.
2197 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2198 bool GuaranteedTailCallOpt) {
2199 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2203 X86TargetLowering::LowerMemArgument(SDValue Chain,
2204 CallingConv::ID CallConv,
2205 const SmallVectorImpl<ISD::InputArg> &Ins,
2206 SDLoc dl, SelectionDAG &DAG,
2207 const CCValAssign &VA,
2208 MachineFrameInfo *MFI,
2210 // Create the nodes corresponding to a load from this parameter slot.
2211 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2212 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2213 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2214 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2217 // If value is passed by pointer we have address passed instead of the value
2219 if (VA.getLocInfo() == CCValAssign::Indirect)
2220 ValVT = VA.getLocVT();
2222 ValVT = VA.getValVT();
2224 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2225 // changed with more analysis.
2226 // In case of tail call optimization mark all arguments mutable. Since they
2227 // could be overwritten by lowering of arguments in case of a tail call.
2228 if (Flags.isByVal()) {
2229 unsigned Bytes = Flags.getByValSize();
2230 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2231 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2232 return DAG.getFrameIndex(FI, getPointerTy());
2234 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2235 VA.getLocMemOffset(), isImmutable);
2236 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2237 return DAG.getLoad(ValVT, dl, Chain, FIN,
2238 MachinePointerInfo::getFixedStack(FI),
2239 false, false, false, 0);
2244 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2245 CallingConv::ID CallConv,
2247 const SmallVectorImpl<ISD::InputArg> &Ins,
2250 SmallVectorImpl<SDValue> &InVals)
2252 MachineFunction &MF = DAG.getMachineFunction();
2253 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2255 const Function* Fn = MF.getFunction();
2256 if (Fn->hasExternalLinkage() &&
2257 Subtarget->isTargetCygMing() &&
2258 Fn->getName() == "main")
2259 FuncInfo->setForceFramePointer(true);
2261 MachineFrameInfo *MFI = MF.getFrameInfo();
2262 bool Is64Bit = Subtarget->is64Bit();
2263 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2265 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2266 "Var args not supported with calling convention fastcc, ghc or hipe");
2268 // Assign locations to all of the incoming arguments.
2269 SmallVector<CCValAssign, 16> ArgLocs;
2270 CCState CCInfo(CallConv, isVarArg, MF, DAG.getTarget(),
2271 ArgLocs, *DAG.getContext());
2273 // Allocate shadow area for Win64
2275 CCInfo.AllocateStack(32, 8);
2277 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2279 unsigned LastVal = ~0U;
2281 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2282 CCValAssign &VA = ArgLocs[i];
2283 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2285 assert(VA.getValNo() != LastVal &&
2286 "Don't support value assigned to multiple locs yet");
2288 LastVal = VA.getValNo();
2290 if (VA.isRegLoc()) {
2291 EVT RegVT = VA.getLocVT();
2292 const TargetRegisterClass *RC;
2293 if (RegVT == MVT::i32)
2294 RC = &X86::GR32RegClass;
2295 else if (Is64Bit && RegVT == MVT::i64)
2296 RC = &X86::GR64RegClass;
2297 else if (RegVT == MVT::f32)
2298 RC = &X86::FR32RegClass;
2299 else if (RegVT == MVT::f64)
2300 RC = &X86::FR64RegClass;
2301 else if (RegVT.is512BitVector())
2302 RC = &X86::VR512RegClass;
2303 else if (RegVT.is256BitVector())
2304 RC = &X86::VR256RegClass;
2305 else if (RegVT.is128BitVector())
2306 RC = &X86::VR128RegClass;
2307 else if (RegVT == MVT::x86mmx)
2308 RC = &X86::VR64RegClass;
2309 else if (RegVT == MVT::i1)
2310 RC = &X86::VK1RegClass;
2311 else if (RegVT == MVT::v8i1)
2312 RC = &X86::VK8RegClass;
2313 else if (RegVT == MVT::v16i1)
2314 RC = &X86::VK16RegClass;
2316 llvm_unreachable("Unknown argument type!");
2318 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2319 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2321 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2322 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2324 if (VA.getLocInfo() == CCValAssign::SExt)
2325 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2326 DAG.getValueType(VA.getValVT()));
2327 else if (VA.getLocInfo() == CCValAssign::ZExt)
2328 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2329 DAG.getValueType(VA.getValVT()));
2330 else if (VA.getLocInfo() == CCValAssign::BCvt)
2331 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2333 if (VA.isExtInLoc()) {
2334 // Handle MMX values passed in XMM regs.
2335 if (RegVT.isVector())
2336 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2338 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2341 assert(VA.isMemLoc());
2342 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2345 // If value is passed via pointer - do a load.
2346 if (VA.getLocInfo() == CCValAssign::Indirect)
2347 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2348 MachinePointerInfo(), false, false, false, 0);
2350 InVals.push_back(ArgValue);
2353 if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) {
2354 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2355 // The x86-64 ABIs require that for returning structs by value we copy
2356 // the sret argument into %rax/%eax (depending on ABI) for the return.
2357 // Win32 requires us to put the sret argument to %eax as well.
2358 // Save the argument into a virtual register so that we can access it
2359 // from the return points.
2360 if (Ins[i].Flags.isSRet()) {
2361 unsigned Reg = FuncInfo->getSRetReturnReg();
2363 MVT PtrTy = getPointerTy();
2364 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2365 FuncInfo->setSRetReturnReg(Reg);
2367 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2368 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2374 unsigned StackSize = CCInfo.getNextStackOffset();
2375 // Align stack specially for tail calls.
2376 if (FuncIsMadeTailCallSafe(CallConv,
2377 MF.getTarget().Options.GuaranteedTailCallOpt))
2378 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2380 // If the function takes variable number of arguments, make a frame index for
2381 // the start of the first vararg value... for expansion of llvm.va_start.
2383 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2384 CallConv != CallingConv::X86_ThisCall)) {
2385 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
2388 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
2390 // FIXME: We should really autogenerate these arrays
2391 static const MCPhysReg GPR64ArgRegsWin64[] = {
2392 X86::RCX, X86::RDX, X86::R8, X86::R9
2394 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2395 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2397 static const MCPhysReg XMMArgRegs64Bit[] = {
2398 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2399 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2401 const MCPhysReg *GPR64ArgRegs;
2402 unsigned NumXMMRegs = 0;
2405 // The XMM registers which might contain var arg parameters are shadowed
2406 // in their paired GPR. So we only need to save the GPR to their home
2408 TotalNumIntRegs = 4;
2409 GPR64ArgRegs = GPR64ArgRegsWin64;
2411 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
2412 GPR64ArgRegs = GPR64ArgRegs64Bit;
2414 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit,
2417 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
2420 bool NoImplicitFloatOps = Fn->getAttributes().
2421 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
2422 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2423 "SSE register cannot be used when SSE is disabled!");
2424 assert(!(NumXMMRegs && MF.getTarget().Options.UseSoftFloat &&
2425 NoImplicitFloatOps) &&
2426 "SSE register cannot be used when SSE is disabled!");
2427 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2428 !Subtarget->hasSSE1())
2429 // Kernel mode asks for SSE to be disabled, so don't push them
2431 TotalNumXMMRegs = 0;
2434 const TargetFrameLowering &TFI = *MF.getTarget().getFrameLowering();
2435 // Get to the caller-allocated home save location. Add 8 to account
2436 // for the return address.
2437 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2438 FuncInfo->setRegSaveFrameIndex(
2439 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2440 // Fixup to set vararg frame on shadow area (4 x i64).
2442 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2444 // For X86-64, if there are vararg parameters that are passed via
2445 // registers, then we must store them to their spots on the stack so
2446 // they may be loaded by deferencing the result of va_next.
2447 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2448 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
2449 FuncInfo->setRegSaveFrameIndex(
2450 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
2454 // Store the integer parameter registers.
2455 SmallVector<SDValue, 8> MemOps;
2456 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2458 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2459 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
2460 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2461 DAG.getIntPtrConstant(Offset));
2462 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
2463 &X86::GR64RegClass);
2464 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2466 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2467 MachinePointerInfo::getFixedStack(
2468 FuncInfo->getRegSaveFrameIndex(), Offset),
2470 MemOps.push_back(Store);
2474 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
2475 // Now store the XMM (fp + vector) parameter registers.
2476 SmallVector<SDValue, 11> SaveXMMOps;
2477 SaveXMMOps.push_back(Chain);
2479 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2480 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
2481 SaveXMMOps.push_back(ALVal);
2483 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2484 FuncInfo->getRegSaveFrameIndex()));
2485 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2486 FuncInfo->getVarArgsFPOffset()));
2488 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
2489 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
2490 &X86::VR128RegClass);
2491 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
2492 SaveXMMOps.push_back(Val);
2494 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2495 MVT::Other, SaveXMMOps));
2498 if (!MemOps.empty())
2499 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2503 // Some CCs need callee pop.
2504 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2505 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2506 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2508 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2509 // If this is an sret function, the return should pop the hidden pointer.
2510 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2511 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2512 argsAreStructReturn(Ins) == StackStructReturn)
2513 FuncInfo->setBytesToPopOnReturn(4);
2517 // RegSaveFrameIndex is X86-64 only.
2518 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2519 if (CallConv == CallingConv::X86_FastCall ||
2520 CallConv == CallingConv::X86_ThisCall)
2521 // fastcc functions can't have varargs.
2522 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2525 FuncInfo->setArgumentStackSize(StackSize);
2531 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2532 SDValue StackPtr, SDValue Arg,
2533 SDLoc dl, SelectionDAG &DAG,
2534 const CCValAssign &VA,
2535 ISD::ArgFlagsTy Flags) const {
2536 unsigned LocMemOffset = VA.getLocMemOffset();
2537 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2538 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2539 if (Flags.isByVal())
2540 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2542 return DAG.getStore(Chain, dl, Arg, PtrOff,
2543 MachinePointerInfo::getStack(LocMemOffset),
2547 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
2548 /// optimization is performed and it is required.
2550 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2551 SDValue &OutRetAddr, SDValue Chain,
2552 bool IsTailCall, bool Is64Bit,
2553 int FPDiff, SDLoc dl) const {
2554 // Adjust the Return address stack slot.
2555 EVT VT = getPointerTy();
2556 OutRetAddr = getReturnAddressFrameIndex(DAG);
2558 // Load the "old" Return address.
2559 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2560 false, false, false, 0);
2561 return SDValue(OutRetAddr.getNode(), 1);
2564 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
2565 /// optimization is performed and it is required (FPDiff!=0).
2566 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2567 SDValue Chain, SDValue RetAddrFrIdx,
2568 EVT PtrVT, unsigned SlotSize,
2569 int FPDiff, SDLoc dl) {
2570 // Store the return address to the appropriate stack slot.
2571 if (!FPDiff) return Chain;
2572 // Calculate the new stack slot for the return address.
2573 int NewReturnAddrFI =
2574 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2576 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2577 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2578 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2584 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2585 SmallVectorImpl<SDValue> &InVals) const {
2586 SelectionDAG &DAG = CLI.DAG;
2588 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2589 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2590 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2591 SDValue Chain = CLI.Chain;
2592 SDValue Callee = CLI.Callee;
2593 CallingConv::ID CallConv = CLI.CallConv;
2594 bool &isTailCall = CLI.IsTailCall;
2595 bool isVarArg = CLI.IsVarArg;
2597 MachineFunction &MF = DAG.getMachineFunction();
2598 bool Is64Bit = Subtarget->is64Bit();
2599 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2600 StructReturnType SR = callIsStructReturn(Outs);
2601 bool IsSibcall = false;
2603 if (MF.getTarget().Options.DisableTailCalls)
2606 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2608 // Force this to be a tail call. The verifier rules are enough to ensure
2609 // that we can lower this successfully without moving the return address
2612 } else if (isTailCall) {
2613 // Check if it's really possible to do a tail call.
2614 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2615 isVarArg, SR != NotStructReturn,
2616 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2617 Outs, OutVals, Ins, DAG);
2619 // Sibcalls are automatically detected tailcalls which do not require
2621 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2628 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2629 "Var args not supported with calling convention fastcc, ghc or hipe");
2631 // Analyze operands of the call, assigning locations to each operand.
2632 SmallVector<CCValAssign, 16> ArgLocs;
2633 CCState CCInfo(CallConv, isVarArg, MF, MF.getTarget(),
2634 ArgLocs, *DAG.getContext());
2636 // Allocate shadow area for Win64
2638 CCInfo.AllocateStack(32, 8);
2640 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2642 // Get a count of how many bytes are to be pushed on the stack.
2643 unsigned NumBytes = CCInfo.getNextStackOffset();
2645 // This is a sibcall. The memory operands are available in caller's
2646 // own caller's stack.
2648 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2649 IsTailCallConvention(CallConv))
2650 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2653 if (isTailCall && !IsSibcall && !IsMustTail) {
2654 // Lower arguments at fp - stackoffset + fpdiff.
2655 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2656 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2658 FPDiff = NumBytesCallerPushed - NumBytes;
2660 // Set the delta of movement of the returnaddr stackslot.
2661 // But only set if delta is greater than previous delta.
2662 if (FPDiff < X86Info->getTCReturnAddrDelta())
2663 X86Info->setTCReturnAddrDelta(FPDiff);
2666 unsigned NumBytesToPush = NumBytes;
2667 unsigned NumBytesToPop = NumBytes;
2669 // If we have an inalloca argument, all stack space has already been allocated
2670 // for us and be right at the top of the stack. We don't support multiple
2671 // arguments passed in memory when using inalloca.
2672 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2674 assert(ArgLocs.back().getLocMemOffset() == 0 &&
2675 "an inalloca argument must be the only memory argument");
2679 Chain = DAG.getCALLSEQ_START(
2680 Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
2682 SDValue RetAddrFrIdx;
2683 // Load return address for tail calls.
2684 if (isTailCall && FPDiff)
2685 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2686 Is64Bit, FPDiff, dl);
2688 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2689 SmallVector<SDValue, 8> MemOpChains;
2692 // Walk the register/memloc assignments, inserting copies/loads. In the case
2693 // of tail call optimization arguments are handle later.
2694 const X86RegisterInfo *RegInfo =
2695 static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
2696 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2697 // Skip inalloca arguments, they have already been written.
2698 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2699 if (Flags.isInAlloca())
2702 CCValAssign &VA = ArgLocs[i];
2703 EVT RegVT = VA.getLocVT();
2704 SDValue Arg = OutVals[i];
2705 bool isByVal = Flags.isByVal();
2707 // Promote the value if needed.
2708 switch (VA.getLocInfo()) {
2709 default: llvm_unreachable("Unknown loc info!");
2710 case CCValAssign::Full: break;
2711 case CCValAssign::SExt:
2712 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2714 case CCValAssign::ZExt:
2715 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2717 case CCValAssign::AExt:
2718 if (RegVT.is128BitVector()) {
2719 // Special case: passing MMX values in XMM registers.
2720 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2721 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2722 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2724 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2726 case CCValAssign::BCvt:
2727 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2729 case CCValAssign::Indirect: {
2730 // Store the argument.
2731 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2732 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2733 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2734 MachinePointerInfo::getFixedStack(FI),
2741 if (VA.isRegLoc()) {
2742 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2743 if (isVarArg && IsWin64) {
2744 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2745 // shadow reg if callee is a varargs function.
2746 unsigned ShadowReg = 0;
2747 switch (VA.getLocReg()) {
2748 case X86::XMM0: ShadowReg = X86::RCX; break;
2749 case X86::XMM1: ShadowReg = X86::RDX; break;
2750 case X86::XMM2: ShadowReg = X86::R8; break;
2751 case X86::XMM3: ShadowReg = X86::R9; break;
2754 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2756 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2757 assert(VA.isMemLoc());
2758 if (!StackPtr.getNode())
2759 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2761 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2762 dl, DAG, VA, Flags));
2766 if (!MemOpChains.empty())
2767 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2769 if (Subtarget->isPICStyleGOT()) {
2770 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2773 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2774 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2776 // If we are tail calling and generating PIC/GOT style code load the
2777 // address of the callee into ECX. The value in ecx is used as target of
2778 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2779 // for tail calls on PIC/GOT architectures. Normally we would just put the
2780 // address of GOT into ebx and then call target@PLT. But for tail calls
2781 // ebx would be restored (since ebx is callee saved) before jumping to the
2784 // Note: The actual moving to ECX is done further down.
2785 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2786 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2787 !G->getGlobal()->hasProtectedVisibility())
2788 Callee = LowerGlobalAddress(Callee, DAG);
2789 else if (isa<ExternalSymbolSDNode>(Callee))
2790 Callee = LowerExternalSymbol(Callee, DAG);
2794 if (Is64Bit && isVarArg && !IsWin64) {
2795 // From AMD64 ABI document:
2796 // For calls that may call functions that use varargs or stdargs
2797 // (prototype-less calls or calls to functions containing ellipsis (...) in
2798 // the declaration) %al is used as hidden argument to specify the number
2799 // of SSE registers used. The contents of %al do not need to match exactly
2800 // the number of registers, but must be an ubound on the number of SSE
2801 // registers used and is in the range 0 - 8 inclusive.
2803 // Count the number of XMM registers allocated.
2804 static const MCPhysReg XMMArgRegs[] = {
2805 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2806 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2808 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2809 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2810 && "SSE registers cannot be used when SSE is disabled");
2812 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2813 DAG.getConstant(NumXMMRegs, MVT::i8)));
2816 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
2817 // don't need this because the eligibility check rejects calls that require
2818 // shuffling arguments passed in memory.
2819 if (!IsSibcall && isTailCall) {
2820 // Force all the incoming stack arguments to be loaded from the stack
2821 // before any new outgoing arguments are stored to the stack, because the
2822 // outgoing stack slots may alias the incoming argument stack slots, and
2823 // the alias isn't otherwise explicit. This is slightly more conservative
2824 // than necessary, because it means that each store effectively depends
2825 // on every argument instead of just those arguments it would clobber.
2826 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2828 SmallVector<SDValue, 8> MemOpChains2;
2831 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2832 CCValAssign &VA = ArgLocs[i];
2835 assert(VA.isMemLoc());
2836 SDValue Arg = OutVals[i];
2837 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2838 // Skip inalloca arguments. They don't require any work.
2839 if (Flags.isInAlloca())
2841 // Create frame index.
2842 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2843 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2844 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2845 FIN = DAG.getFrameIndex(FI, getPointerTy());
2847 if (Flags.isByVal()) {
2848 // Copy relative to framepointer.
2849 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2850 if (!StackPtr.getNode())
2851 StackPtr = DAG.getCopyFromReg(Chain, dl,
2852 RegInfo->getStackRegister(),
2854 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2856 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2860 // Store relative to framepointer.
2861 MemOpChains2.push_back(
2862 DAG.getStore(ArgChain, dl, Arg, FIN,
2863 MachinePointerInfo::getFixedStack(FI),
2868 if (!MemOpChains2.empty())
2869 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
2871 // Store the return address to the appropriate stack slot.
2872 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
2873 getPointerTy(), RegInfo->getSlotSize(),
2877 // Build a sequence of copy-to-reg nodes chained together with token chain
2878 // and flag operands which copy the outgoing args into registers.
2880 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2881 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2882 RegsToPass[i].second, InFlag);
2883 InFlag = Chain.getValue(1);
2886 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
2887 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2888 // In the 64-bit large code model, we have to make all calls
2889 // through a register, since the call instruction's 32-bit
2890 // pc-relative offset may not be large enough to hold the whole
2892 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2893 // If the callee is a GlobalAddress node (quite common, every direct call
2894 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2897 // We should use extra load for direct calls to dllimported functions in
2899 const GlobalValue *GV = G->getGlobal();
2900 if (!GV->hasDLLImportStorageClass()) {
2901 unsigned char OpFlags = 0;
2902 bool ExtraLoad = false;
2903 unsigned WrapperKind = ISD::DELETED_NODE;
2905 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2906 // external symbols most go through the PLT in PIC mode. If the symbol
2907 // has hidden or protected visibility, or if it is static or local, then
2908 // we don't need to use the PLT - we can directly call it.
2909 if (Subtarget->isTargetELF() &&
2910 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
2911 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2912 OpFlags = X86II::MO_PLT;
2913 } else if (Subtarget->isPICStyleStubAny() &&
2914 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2915 (!Subtarget->getTargetTriple().isMacOSX() ||
2916 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2917 // PC-relative references to external symbols should go through $stub,
2918 // unless we're building with the leopard linker or later, which
2919 // automatically synthesizes these stubs.
2920 OpFlags = X86II::MO_DARWIN_STUB;
2921 } else if (Subtarget->isPICStyleRIPRel() &&
2922 isa<Function>(GV) &&
2923 cast<Function>(GV)->getAttributes().
2924 hasAttribute(AttributeSet::FunctionIndex,
2925 Attribute::NonLazyBind)) {
2926 // If the function is marked as non-lazy, generate an indirect call
2927 // which loads from the GOT directly. This avoids runtime overhead
2928 // at the cost of eager binding (and one extra byte of encoding).
2929 OpFlags = X86II::MO_GOTPCREL;
2930 WrapperKind = X86ISD::WrapperRIP;
2934 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2935 G->getOffset(), OpFlags);
2937 // Add a wrapper if needed.
2938 if (WrapperKind != ISD::DELETED_NODE)
2939 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2940 // Add extra indirection if needed.
2942 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2943 MachinePointerInfo::getGOT(),
2944 false, false, false, 0);
2946 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2947 unsigned char OpFlags = 0;
2949 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2950 // external symbols should go through the PLT.
2951 if (Subtarget->isTargetELF() &&
2952 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
2953 OpFlags = X86II::MO_PLT;
2954 } else if (Subtarget->isPICStyleStubAny() &&
2955 (!Subtarget->getTargetTriple().isMacOSX() ||
2956 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2957 // PC-relative references to external symbols should go through $stub,
2958 // unless we're building with the leopard linker or later, which
2959 // automatically synthesizes these stubs.
2960 OpFlags = X86II::MO_DARWIN_STUB;
2963 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2967 // Returns a chain & a flag for retval copy to use.
2968 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2969 SmallVector<SDValue, 8> Ops;
2971 if (!IsSibcall && isTailCall) {
2972 Chain = DAG.getCALLSEQ_END(Chain,
2973 DAG.getIntPtrConstant(NumBytesToPop, true),
2974 DAG.getIntPtrConstant(0, true), InFlag, dl);
2975 InFlag = Chain.getValue(1);
2978 Ops.push_back(Chain);
2979 Ops.push_back(Callee);
2982 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2984 // Add argument registers to the end of the list so that they are known live
2986 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2987 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2988 RegsToPass[i].second.getValueType()));
2990 // Add a register mask operand representing the call-preserved registers.
2991 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
2992 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2993 assert(Mask && "Missing call preserved mask for calling convention");
2994 Ops.push_back(DAG.getRegisterMask(Mask));
2996 if (InFlag.getNode())
2997 Ops.push_back(InFlag);
3001 //// If this is the first return lowered for this function, add the regs
3002 //// to the liveout set for the function.
3003 // This isn't right, although it's probably harmless on x86; liveouts
3004 // should be computed from returns not tail calls. Consider a void
3005 // function making a tail call to a function returning int.
3006 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3009 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3010 InFlag = Chain.getValue(1);
3012 // Create the CALLSEQ_END node.
3013 unsigned NumBytesForCalleeToPop;
3014 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3015 DAG.getTarget().Options.GuaranteedTailCallOpt))
3016 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3017 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3018 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3019 SR == StackStructReturn)
3020 // If this is a call to a struct-return function, the callee
3021 // pops the hidden struct pointer, so we have to push it back.
3022 // This is common for Darwin/X86, Linux & Mingw32 targets.
3023 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3024 NumBytesForCalleeToPop = 4;
3026 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3028 // Returns a flag for retval copy to use.
3030 Chain = DAG.getCALLSEQ_END(Chain,
3031 DAG.getIntPtrConstant(NumBytesToPop, true),
3032 DAG.getIntPtrConstant(NumBytesForCalleeToPop,
3035 InFlag = Chain.getValue(1);
3038 // Handle result values, copying them out of physregs into vregs that we
3040 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3041 Ins, dl, DAG, InVals);
3044 //===----------------------------------------------------------------------===//
3045 // Fast Calling Convention (tail call) implementation
3046 //===----------------------------------------------------------------------===//
3048 // Like std call, callee cleans arguments, convention except that ECX is
3049 // reserved for storing the tail called function address. Only 2 registers are
3050 // free for argument passing (inreg). Tail call optimization is performed
3052 // * tailcallopt is enabled
3053 // * caller/callee are fastcc
3054 // On X86_64 architecture with GOT-style position independent code only local
3055 // (within module) calls are supported at the moment.
3056 // To keep the stack aligned according to platform abi the function
3057 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3058 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3059 // If a tail called function callee has more arguments than the caller the
3060 // caller needs to make sure that there is room to move the RETADDR to. This is
3061 // achieved by reserving an area the size of the argument delta right after the
3062 // original RETADDR, but before the saved framepointer or the spilled registers
3063 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3075 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3076 /// for a 16 byte align requirement.
3078 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3079 SelectionDAG& DAG) const {
3080 MachineFunction &MF = DAG.getMachineFunction();
3081 const TargetMachine &TM = MF.getTarget();
3082 const X86RegisterInfo *RegInfo =
3083 static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
3084 const TargetFrameLowering &TFI = *TM.getFrameLowering();
3085 unsigned StackAlignment = TFI.getStackAlignment();
3086 uint64_t AlignMask = StackAlignment - 1;
3087 int64_t Offset = StackSize;
3088 unsigned SlotSize = RegInfo->getSlotSize();
3089 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3090 // Number smaller than 12 so just add the difference.
3091 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3093 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3094 Offset = ((~AlignMask) & Offset) + StackAlignment +
3095 (StackAlignment-SlotSize);
3100 /// MatchingStackOffset - Return true if the given stack call argument is
3101 /// already available in the same position (relatively) of the caller's
3102 /// incoming argument stack.
3104 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3105 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3106 const X86InstrInfo *TII) {
3107 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3109 if (Arg.getOpcode() == ISD::CopyFromReg) {
3110 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3111 if (!TargetRegisterInfo::isVirtualRegister(VR))
3113 MachineInstr *Def = MRI->getVRegDef(VR);
3116 if (!Flags.isByVal()) {
3117 if (!TII->isLoadFromStackSlot(Def, FI))
3120 unsigned Opcode = Def->getOpcode();
3121 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
3122 Def->getOperand(1).isFI()) {
3123 FI = Def->getOperand(1).getIndex();
3124 Bytes = Flags.getByValSize();
3128 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3129 if (Flags.isByVal())
3130 // ByVal argument is passed in as a pointer but it's now being
3131 // dereferenced. e.g.
3132 // define @foo(%struct.X* %A) {
3133 // tail call @bar(%struct.X* byval %A)
3136 SDValue Ptr = Ld->getBasePtr();
3137 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3140 FI = FINode->getIndex();
3141 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3142 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3143 FI = FINode->getIndex();
3144 Bytes = Flags.getByValSize();
3148 assert(FI != INT_MAX);
3149 if (!MFI->isFixedObjectIndex(FI))
3151 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3154 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3155 /// for tail call optimization. Targets which want to do tail call
3156 /// optimization should implement this function.
3158 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3159 CallingConv::ID CalleeCC,
3161 bool isCalleeStructRet,
3162 bool isCallerStructRet,
3164 const SmallVectorImpl<ISD::OutputArg> &Outs,
3165 const SmallVectorImpl<SDValue> &OutVals,
3166 const SmallVectorImpl<ISD::InputArg> &Ins,
3167 SelectionDAG &DAG) const {
3168 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3171 // If -tailcallopt is specified, make fastcc functions tail-callable.
3172 const MachineFunction &MF = DAG.getMachineFunction();
3173 const Function *CallerF = MF.getFunction();
3175 // If the function return type is x86_fp80 and the callee return type is not,
3176 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3177 // perform a tailcall optimization here.
3178 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3181 CallingConv::ID CallerCC = CallerF->getCallingConv();
3182 bool CCMatch = CallerCC == CalleeCC;
3183 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3184 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3186 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3187 if (IsTailCallConvention(CalleeCC) && CCMatch)
3192 // Look for obvious safe cases to perform tail call optimization that do not
3193 // require ABI changes. This is what gcc calls sibcall.
3195 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3196 // emit a special epilogue.
3197 const X86RegisterInfo *RegInfo =
3198 static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
3199 if (RegInfo->needsStackRealignment(MF))
3202 // Also avoid sibcall optimization if either caller or callee uses struct
3203 // return semantics.
3204 if (isCalleeStructRet || isCallerStructRet)
3207 // An stdcall/thiscall caller is expected to clean up its arguments; the
3208 // callee isn't going to do that.
3209 // FIXME: this is more restrictive than needed. We could produce a tailcall
3210 // when the stack adjustment matches. For example, with a thiscall that takes
3211 // only one argument.
3212 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3213 CallerCC == CallingConv::X86_ThisCall))
3216 // Do not sibcall optimize vararg calls unless all arguments are passed via
3218 if (isVarArg && !Outs.empty()) {
3220 // Optimizing for varargs on Win64 is unlikely to be safe without
3221 // additional testing.
3222 if (IsCalleeWin64 || IsCallerWin64)
3225 SmallVector<CCValAssign, 16> ArgLocs;
3226 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
3227 DAG.getTarget(), ArgLocs, *DAG.getContext());
3229 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3230 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3231 if (!ArgLocs[i].isRegLoc())
3235 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3236 // stack. Therefore, if it's not used by the call it is not safe to optimize
3237 // this into a sibcall.
3238 bool Unused = false;
3239 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3246 SmallVector<CCValAssign, 16> RVLocs;
3247 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
3248 DAG.getTarget(), RVLocs, *DAG.getContext());
3249 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3250 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3251 CCValAssign &VA = RVLocs[i];
3252 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
3257 // If the calling conventions do not match, then we'd better make sure the
3258 // results are returned in the same way as what the caller expects.
3260 SmallVector<CCValAssign, 16> RVLocs1;
3261 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
3262 DAG.getTarget(), RVLocs1, *DAG.getContext());
3263 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3265 SmallVector<CCValAssign, 16> RVLocs2;
3266 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
3267 DAG.getTarget(), RVLocs2, *DAG.getContext());
3268 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3270 if (RVLocs1.size() != RVLocs2.size())
3272 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3273 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3275 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3277 if (RVLocs1[i].isRegLoc()) {
3278 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3281 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3287 // If the callee takes no arguments then go on to check the results of the
3289 if (!Outs.empty()) {
3290 // Check if stack adjustment is needed. For now, do not do this if any
3291 // argument is passed on the stack.
3292 SmallVector<CCValAssign, 16> ArgLocs;
3293 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
3294 DAG.getTarget(), ArgLocs, *DAG.getContext());
3296 // Allocate shadow area for Win64
3298 CCInfo.AllocateStack(32, 8);
3300 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3301 if (CCInfo.getNextStackOffset()) {
3302 MachineFunction &MF = DAG.getMachineFunction();
3303 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3306 // Check if the arguments are already laid out in the right way as
3307 // the caller's fixed stack objects.
3308 MachineFrameInfo *MFI = MF.getFrameInfo();
3309 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3310 const X86InstrInfo *TII =
3311 static_cast<const X86InstrInfo *>(DAG.getTarget().getInstrInfo());
3312 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3313 CCValAssign &VA = ArgLocs[i];
3314 SDValue Arg = OutVals[i];
3315 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3316 if (VA.getLocInfo() == CCValAssign::Indirect)
3318 if (!VA.isRegLoc()) {
3319 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3326 // If the tailcall address may be in a register, then make sure it's
3327 // possible to register allocate for it. In 32-bit, the call address can
3328 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3329 // callee-saved registers are restored. These happen to be the same
3330 // registers used to pass 'inreg' arguments so watch out for those.
3331 if (!Subtarget->is64Bit() &&
3332 ((!isa<GlobalAddressSDNode>(Callee) &&
3333 !isa<ExternalSymbolSDNode>(Callee)) ||
3334 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3335 unsigned NumInRegs = 0;
3336 // In PIC we need an extra register to formulate the address computation
3338 unsigned MaxInRegs =
3339 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3341 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3342 CCValAssign &VA = ArgLocs[i];
3345 unsigned Reg = VA.getLocReg();
3348 case X86::EAX: case X86::EDX: case X86::ECX:
3349 if (++NumInRegs == MaxInRegs)
3361 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3362 const TargetLibraryInfo *libInfo) const {
3363 return X86::createFastISel(funcInfo, libInfo);
3366 //===----------------------------------------------------------------------===//
3367 // Other Lowering Hooks
3368 //===----------------------------------------------------------------------===//
3370 static bool MayFoldLoad(SDValue Op) {
3371 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3374 static bool MayFoldIntoStore(SDValue Op) {
3375 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3378 static bool isTargetShuffle(unsigned Opcode) {
3380 default: return false;
3381 case X86ISD::PSHUFD:
3382 case X86ISD::PSHUFHW:
3383 case X86ISD::PSHUFLW:
3385 case X86ISD::PALIGNR:
3386 case X86ISD::MOVLHPS:
3387 case X86ISD::MOVLHPD:
3388 case X86ISD::MOVHLPS:
3389 case X86ISD::MOVLPS:
3390 case X86ISD::MOVLPD:
3391 case X86ISD::MOVSHDUP:
3392 case X86ISD::MOVSLDUP:
3393 case X86ISD::MOVDDUP:
3396 case X86ISD::UNPCKL:
3397 case X86ISD::UNPCKH:
3398 case X86ISD::VPERMILP:
3399 case X86ISD::VPERM2X128:
3400 case X86ISD::VPERMI:
3405 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3406 SDValue V1, SelectionDAG &DAG) {
3408 default: llvm_unreachable("Unknown x86 shuffle node");
3409 case X86ISD::MOVSHDUP:
3410 case X86ISD::MOVSLDUP:
3411 case X86ISD::MOVDDUP:
3412 return DAG.getNode(Opc, dl, VT, V1);
3416 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3417 SDValue V1, unsigned TargetMask,
3418 SelectionDAG &DAG) {
3420 default: llvm_unreachable("Unknown x86 shuffle node");
3421 case X86ISD::PSHUFD:
3422 case X86ISD::PSHUFHW:
3423 case X86ISD::PSHUFLW:
3424 case X86ISD::VPERMILP:
3425 case X86ISD::VPERMI:
3426 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3430 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3431 SDValue V1, SDValue V2, unsigned TargetMask,
3432 SelectionDAG &DAG) {
3434 default: llvm_unreachable("Unknown x86 shuffle node");
3435 case X86ISD::PALIGNR:
3437 case X86ISD::VPERM2X128:
3438 return DAG.getNode(Opc, dl, VT, V1, V2,
3439 DAG.getConstant(TargetMask, MVT::i8));
3443 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3444 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3446 default: llvm_unreachable("Unknown x86 shuffle node");
3447 case X86ISD::MOVLHPS:
3448 case X86ISD::MOVLHPD:
3449 case X86ISD::MOVHLPS:
3450 case X86ISD::MOVLPS:
3451 case X86ISD::MOVLPD:
3454 case X86ISD::UNPCKL:
3455 case X86ISD::UNPCKH:
3456 return DAG.getNode(Opc, dl, VT, V1, V2);
3460 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3461 MachineFunction &MF = DAG.getMachineFunction();
3462 const X86RegisterInfo *RegInfo =
3463 static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
3464 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3465 int ReturnAddrIndex = FuncInfo->getRAIndex();
3467 if (ReturnAddrIndex == 0) {
3468 // Set up a frame object for the return address.
3469 unsigned SlotSize = RegInfo->getSlotSize();
3470 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3473 FuncInfo->setRAIndex(ReturnAddrIndex);
3476 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3479 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3480 bool hasSymbolicDisplacement) {
3481 // Offset should fit into 32 bit immediate field.
3482 if (!isInt<32>(Offset))
3485 // If we don't have a symbolic displacement - we don't have any extra
3487 if (!hasSymbolicDisplacement)
3490 // FIXME: Some tweaks might be needed for medium code model.
3491 if (M != CodeModel::Small && M != CodeModel::Kernel)
3494 // For small code model we assume that latest object is 16MB before end of 31
3495 // bits boundary. We may also accept pretty large negative constants knowing
3496 // that all objects are in the positive half of address space.
3497 if (M == CodeModel::Small && Offset < 16*1024*1024)
3500 // For kernel code model we know that all object resist in the negative half
3501 // of 32bits address space. We may not accept negative offsets, since they may
3502 // be just off and we may accept pretty large positive ones.
3503 if (M == CodeModel::Kernel && Offset > 0)
3509 /// isCalleePop - Determines whether the callee is required to pop its
3510 /// own arguments. Callee pop is necessary to support tail calls.
3511 bool X86::isCalleePop(CallingConv::ID CallingConv,
3512 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3516 switch (CallingConv) {
3519 case CallingConv::X86_StdCall:
3521 case CallingConv::X86_FastCall:
3523 case CallingConv::X86_ThisCall:
3525 case CallingConv::Fast:
3527 case CallingConv::GHC:
3529 case CallingConv::HiPE:
3534 /// \brief Return true if the condition is an unsigned comparison operation.
3535 static bool isX86CCUnsigned(unsigned X86CC) {
3537 default: llvm_unreachable("Invalid integer condition!");
3538 case X86::COND_E: return true;
3539 case X86::COND_G: return false;
3540 case X86::COND_GE: return false;
3541 case X86::COND_L: return false;
3542 case X86::COND_LE: return false;
3543 case X86::COND_NE: return true;
3544 case X86::COND_B: return true;
3545 case X86::COND_A: return true;
3546 case X86::COND_BE: return true;
3547 case X86::COND_AE: return true;
3549 llvm_unreachable("covered switch fell through?!");
3552 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3553 /// specific condition code, returning the condition code and the LHS/RHS of the
3554 /// comparison to make.
3555 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3556 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3558 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3559 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3560 // X > -1 -> X == 0, jump !sign.
3561 RHS = DAG.getConstant(0, RHS.getValueType());
3562 return X86::COND_NS;
3564 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3565 // X < 0 -> X == 0, jump on sign.
3568 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3570 RHS = DAG.getConstant(0, RHS.getValueType());
3571 return X86::COND_LE;
3575 switch (SetCCOpcode) {
3576 default: llvm_unreachable("Invalid integer condition!");
3577 case ISD::SETEQ: return X86::COND_E;
3578 case ISD::SETGT: return X86::COND_G;
3579 case ISD::SETGE: return X86::COND_GE;
3580 case ISD::SETLT: return X86::COND_L;
3581 case ISD::SETLE: return X86::COND_LE;
3582 case ISD::SETNE: return X86::COND_NE;
3583 case ISD::SETULT: return X86::COND_B;
3584 case ISD::SETUGT: return X86::COND_A;
3585 case ISD::SETULE: return X86::COND_BE;
3586 case ISD::SETUGE: return X86::COND_AE;
3590 // First determine if it is required or is profitable to flip the operands.
3592 // If LHS is a foldable load, but RHS is not, flip the condition.
3593 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3594 !ISD::isNON_EXTLoad(RHS.getNode())) {
3595 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3596 std::swap(LHS, RHS);
3599 switch (SetCCOpcode) {
3605 std::swap(LHS, RHS);
3609 // On a floating point condition, the flags are set as follows:
3611 // 0 | 0 | 0 | X > Y
3612 // 0 | 0 | 1 | X < Y
3613 // 1 | 0 | 0 | X == Y
3614 // 1 | 1 | 1 | unordered
3615 switch (SetCCOpcode) {
3616 default: llvm_unreachable("Condcode should be pre-legalized away");
3618 case ISD::SETEQ: return X86::COND_E;
3619 case ISD::SETOLT: // flipped
3621 case ISD::SETGT: return X86::COND_A;
3622 case ISD::SETOLE: // flipped
3624 case ISD::SETGE: return X86::COND_AE;
3625 case ISD::SETUGT: // flipped
3627 case ISD::SETLT: return X86::COND_B;
3628 case ISD::SETUGE: // flipped
3630 case ISD::SETLE: return X86::COND_BE;
3632 case ISD::SETNE: return X86::COND_NE;
3633 case ISD::SETUO: return X86::COND_P;
3634 case ISD::SETO: return X86::COND_NP;
3636 case ISD::SETUNE: return X86::COND_INVALID;
3640 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3641 /// code. Current x86 isa includes the following FP cmov instructions:
3642 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3643 static bool hasFPCMov(unsigned X86CC) {
3659 /// isFPImmLegal - Returns true if the target can instruction select the
3660 /// specified FP immediate natively. If false, the legalizer will
3661 /// materialize the FP immediate as a load from a constant pool.
3662 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3663 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3664 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3670 /// \brief Returns true if it is beneficial to convert a load of a constant
3671 /// to just the constant itself.
3672 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3674 assert(Ty->isIntegerTy());
3676 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3677 if (BitSize == 0 || BitSize > 64)
3682 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3683 /// the specified range (L, H].
3684 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3685 return (Val < 0) || (Val >= Low && Val < Hi);
3688 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3689 /// specified value.
3690 static bool isUndefOrEqual(int Val, int CmpVal) {
3691 return (Val < 0 || Val == CmpVal);
3694 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3695 /// from position Pos and ending in Pos+Size, falls within the specified
3696 /// sequential range (L, L+Pos]. or is undef.
3697 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3698 unsigned Pos, unsigned Size, int Low) {
3699 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3700 if (!isUndefOrEqual(Mask[i], Low))
3705 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3706 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3707 /// the second operand.
3708 static bool isPSHUFDMask(ArrayRef<int> Mask, MVT VT) {
3709 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3710 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3711 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3712 return (Mask[0] < 2 && Mask[1] < 2);
3716 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3717 /// is suitable for input to PSHUFHW.
3718 static bool isPSHUFHWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3719 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3722 // Lower quadword copied in order or undef.
3723 if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
3726 // Upper quadword shuffled.
3727 for (unsigned i = 4; i != 8; ++i)
3728 if (!isUndefOrInRange(Mask[i], 4, 8))
3731 if (VT == MVT::v16i16) {
3732 // Lower quadword copied in order or undef.
3733 if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
3736 // Upper quadword shuffled.
3737 for (unsigned i = 12; i != 16; ++i)
3738 if (!isUndefOrInRange(Mask[i], 12, 16))
3745 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3746 /// is suitable for input to PSHUFLW.
3747 static bool isPSHUFLWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3748 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3751 // Upper quadword copied in order.
3752 if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
3755 // Lower quadword shuffled.
3756 for (unsigned i = 0; i != 4; ++i)
3757 if (!isUndefOrInRange(Mask[i], 0, 4))
3760 if (VT == MVT::v16i16) {
3761 // Upper quadword copied in order.
3762 if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
3765 // Lower quadword shuffled.
3766 for (unsigned i = 8; i != 12; ++i)
3767 if (!isUndefOrInRange(Mask[i], 8, 12))
3774 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3775 /// is suitable for input to PALIGNR.
3776 static bool isPALIGNRMask(ArrayRef<int> Mask, MVT VT,
3777 const X86Subtarget *Subtarget) {
3778 if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
3779 (VT.is256BitVector() && !Subtarget->hasInt256()))
3782 unsigned NumElts = VT.getVectorNumElements();
3783 unsigned NumLanes = VT.is512BitVector() ? 1: VT.getSizeInBits()/128;
3784 unsigned NumLaneElts = NumElts/NumLanes;
3786 // Do not handle 64-bit element shuffles with palignr.
3787 if (NumLaneElts == 2)
3790 for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
3792 for (i = 0; i != NumLaneElts; ++i) {
3797 // Lane is all undef, go to next lane
3798 if (i == NumLaneElts)
3801 int Start = Mask[i+l];
3803 // Make sure its in this lane in one of the sources
3804 if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
3805 !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
3808 // If not lane 0, then we must match lane 0
3809 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
3812 // Correct second source to be contiguous with first source
3813 if (Start >= (int)NumElts)
3814 Start -= NumElts - NumLaneElts;
3816 // Make sure we're shifting in the right direction.
3817 if (Start <= (int)(i+l))
3822 // Check the rest of the elements to see if they are consecutive.
3823 for (++i; i != NumLaneElts; ++i) {
3824 int Idx = Mask[i+l];
3826 // Make sure its in this lane
3827 if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
3828 !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
3831 // If not lane 0, then we must match lane 0
3832 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
3835 if (Idx >= (int)NumElts)
3836 Idx -= NumElts - NumLaneElts;
3838 if (!isUndefOrEqual(Idx, Start+i))
3847 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3848 /// the two vector operands have swapped position.
3849 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
3850 unsigned NumElems) {
3851 for (unsigned i = 0; i != NumElems; ++i) {
3855 else if (idx < (int)NumElems)
3856 Mask[i] = idx + NumElems;
3858 Mask[i] = idx - NumElems;
3862 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3863 /// specifies a shuffle of elements that is suitable for input to 128/256-bit
3864 /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
3865 /// reverse of what x86 shuffles want.
3866 static bool isSHUFPMask(ArrayRef<int> Mask, MVT VT, bool Commuted = false) {
3868 unsigned NumElems = VT.getVectorNumElements();
3869 unsigned NumLanes = VT.getSizeInBits()/128;
3870 unsigned NumLaneElems = NumElems/NumLanes;
3872 if (NumLaneElems != 2 && NumLaneElems != 4)
3875 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3876 bool symetricMaskRequired =
3877 (VT.getSizeInBits() >= 256) && (EltSize == 32);
3879 // VSHUFPSY divides the resulting vector into 4 chunks.
3880 // The sources are also splitted into 4 chunks, and each destination
3881 // chunk must come from a different source chunk.
3883 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
3884 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
3886 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
3887 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
3889 // VSHUFPDY divides the resulting vector into 4 chunks.
3890 // The sources are also splitted into 4 chunks, and each destination
3891 // chunk must come from a different source chunk.
3893 // SRC1 => X3 X2 X1 X0
3894 // SRC2 => Y3 Y2 Y1 Y0
3896 // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
3898 SmallVector<int, 4> MaskVal(NumLaneElems, -1);
3899 unsigned HalfLaneElems = NumLaneElems/2;
3900 for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
3901 for (unsigned i = 0; i != NumLaneElems; ++i) {
3902 int Idx = Mask[i+l];
3903 unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
3904 if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
3906 // For VSHUFPSY, the mask of the second half must be the same as the
3907 // first but with the appropriate offsets. This works in the same way as
3908 // VPERMILPS works with masks.
3909 if (!symetricMaskRequired || Idx < 0)
3911 if (MaskVal[i] < 0) {
3912 MaskVal[i] = Idx - l;
3915 if ((signed)(Idx - l) != MaskVal[i])
3923 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3924 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3925 static bool isMOVHLPSMask(ArrayRef<int> Mask, MVT VT) {
3926 if (!VT.is128BitVector())
3929 unsigned NumElems = VT.getVectorNumElements();
3934 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3935 return isUndefOrEqual(Mask[0], 6) &&
3936 isUndefOrEqual(Mask[1], 7) &&
3937 isUndefOrEqual(Mask[2], 2) &&
3938 isUndefOrEqual(Mask[3], 3);
3941 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3942 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3944 static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, MVT VT) {
3945 if (!VT.is128BitVector())
3948 unsigned NumElems = VT.getVectorNumElements();
3953 return isUndefOrEqual(Mask[0], 2) &&
3954 isUndefOrEqual(Mask[1], 3) &&
3955 isUndefOrEqual(Mask[2], 2) &&
3956 isUndefOrEqual(Mask[3], 3);
3959 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3960 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3961 static bool isMOVLPMask(ArrayRef<int> Mask, MVT VT) {
3962 if (!VT.is128BitVector())
3965 unsigned NumElems = VT.getVectorNumElements();
3967 if (NumElems != 2 && NumElems != 4)
3970 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3971 if (!isUndefOrEqual(Mask[i], i + NumElems))
3974 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
3975 if (!isUndefOrEqual(Mask[i], i))
3981 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3982 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3983 static bool isMOVLHPSMask(ArrayRef<int> Mask, MVT VT) {
3984 if (!VT.is128BitVector())
3987 unsigned NumElems = VT.getVectorNumElements();
3989 if (NumElems != 2 && NumElems != 4)
3992 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3993 if (!isUndefOrEqual(Mask[i], i))
3996 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3997 if (!isUndefOrEqual(Mask[i + e], i + NumElems))
4003 /// isINSERTPSMask - Return true if the specified VECTOR_SHUFFLE operand
4004 /// specifies a shuffle of elements that is suitable for input to INSERTPS.
4005 /// i. e: If all but one element come from the same vector.
4006 static bool isINSERTPSMask(ArrayRef<int> Mask, MVT VT) {
4007 // TODO: Deal with AVX's VINSERTPS
4008 if (!VT.is128BitVector() || (VT != MVT::v4f32 && VT != MVT::v4i32))
4011 unsigned CorrectPosV1 = 0;
4012 unsigned CorrectPosV2 = 0;
4013 for (int i = 0, e = (int)VT.getVectorNumElements(); i != e; ++i) {
4014 if (Mask[i] == -1) {
4022 else if (Mask[i] == i + 4)
4026 if (CorrectPosV1 == 3 || CorrectPosV2 == 3)
4027 // We have 3 elements (undefs count as elements from any vector) from one
4028 // vector, and one from another.
4035 // Some special combinations that can be optimized.
4038 SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
4039 SelectionDAG &DAG) {
4040 MVT VT = SVOp->getSimpleValueType(0);
4043 if (VT != MVT::v8i32 && VT != MVT::v8f32)
4046 ArrayRef<int> Mask = SVOp->getMask();
4048 // These are the special masks that may be optimized.
4049 static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
4050 static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
4051 bool MatchEvenMask = true;
4052 bool MatchOddMask = true;
4053 for (int i=0; i<8; ++i) {
4054 if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
4055 MatchEvenMask = false;
4056 if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
4057 MatchOddMask = false;
4060 if (!MatchEvenMask && !MatchOddMask)
4063 SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
4065 SDValue Op0 = SVOp->getOperand(0);
4066 SDValue Op1 = SVOp->getOperand(1);
4068 if (MatchEvenMask) {
4069 // Shift the second operand right to 32 bits.
4070 static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
4071 Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
4073 // Shift the first operand left to 32 bits.
4074 static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
4075 Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
4077 static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
4078 return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
4081 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
4082 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
4083 static bool isUNPCKLMask(ArrayRef<int> Mask, MVT VT,
4084 bool HasInt256, bool V2IsSplat = false) {
4086 assert(VT.getSizeInBits() >= 128 &&
4087 "Unsupported vector type for unpckl");
4089 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4091 unsigned NumOf256BitLanes;
4092 unsigned NumElts = VT.getVectorNumElements();
4093 if (VT.is256BitVector()) {
4094 if (NumElts != 4 && NumElts != 8 &&
4095 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4098 NumOf256BitLanes = 1;
4099 } else if (VT.is512BitVector()) {
4100 assert(VT.getScalarType().getSizeInBits() >= 32 &&
4101 "Unsupported vector type for unpckh");
4103 NumOf256BitLanes = 2;
4106 NumOf256BitLanes = 1;
4109 unsigned NumEltsInStride = NumElts/NumOf256BitLanes;
4110 unsigned NumLaneElts = NumEltsInStride/NumLanes;
4112 for (unsigned l256 = 0; l256 < NumOf256BitLanes; l256 += 1) {
4113 for (unsigned l = 0; l != NumEltsInStride; l += NumLaneElts) {
4114 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4115 int BitI = Mask[l256*NumEltsInStride+l+i];
4116 int BitI1 = Mask[l256*NumEltsInStride+l+i+1];
4117 if (!isUndefOrEqual(BitI, j+l256*NumElts))
4119 if (V2IsSplat && !isUndefOrEqual(BitI1, NumElts))
4121 if (!isUndefOrEqual(BitI1, j+l256*NumElts+NumEltsInStride))
4129 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
4130 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
4131 static bool isUNPCKHMask(ArrayRef<int> Mask, MVT VT,
4132 bool HasInt256, bool V2IsSplat = false) {
4133 assert(VT.getSizeInBits() >= 128 &&
4134 "Unsupported vector type for unpckh");
4136 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4138 unsigned NumOf256BitLanes;
4139 unsigned NumElts = VT.getVectorNumElements();
4140 if (VT.is256BitVector()) {
4141 if (NumElts != 4 && NumElts != 8 &&
4142 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4145 NumOf256BitLanes = 1;
4146 } else if (VT.is512BitVector()) {
4147 assert(VT.getScalarType().getSizeInBits() >= 32 &&
4148 "Unsupported vector type for unpckh");
4150 NumOf256BitLanes = 2;
4153 NumOf256BitLanes = 1;
4156 unsigned NumEltsInStride = NumElts/NumOf256BitLanes;
4157 unsigned NumLaneElts = NumEltsInStride/NumLanes;
4159 for (unsigned l256 = 0; l256 < NumOf256BitLanes; l256 += 1) {
4160 for (unsigned l = 0; l != NumEltsInStride; l += NumLaneElts) {
4161 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4162 int BitI = Mask[l256*NumEltsInStride+l+i];
4163 int BitI1 = Mask[l256*NumEltsInStride+l+i+1];
4164 if (!isUndefOrEqual(BitI, j+l256*NumElts))
4166 if (V2IsSplat && !isUndefOrEqual(BitI1, NumElts))
4168 if (!isUndefOrEqual(BitI1, j+l256*NumElts+NumEltsInStride))
4176 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
4177 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
4179 static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4180 unsigned NumElts = VT.getVectorNumElements();
4181 bool Is256BitVec = VT.is256BitVector();
4183 if (VT.is512BitVector())
4185 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4186 "Unsupported vector type for unpckh");
4188 if (Is256BitVec && NumElts != 4 && NumElts != 8 &&
4189 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4192 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
4193 // FIXME: Need a better way to get rid of this, there's no latency difference
4194 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
4195 // the former later. We should also remove the "_undef" special mask.
4196 if (NumElts == 4 && Is256BitVec)
4199 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4200 // independently on 128-bit lanes.
4201 unsigned NumLanes = VT.getSizeInBits()/128;
4202 unsigned NumLaneElts = NumElts/NumLanes;
4204 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4205 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4206 int BitI = Mask[l+i];
4207 int BitI1 = Mask[l+i+1];
4209 if (!isUndefOrEqual(BitI, j))
4211 if (!isUndefOrEqual(BitI1, j))
4219 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
4220 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
4222 static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4223 unsigned NumElts = VT.getVectorNumElements();
4225 if (VT.is512BitVector())
4228 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4229 "Unsupported vector type for unpckh");
4231 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4232 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4235 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4236 // independently on 128-bit lanes.
4237 unsigned NumLanes = VT.getSizeInBits()/128;
4238 unsigned NumLaneElts = NumElts/NumLanes;
4240 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4241 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4242 int BitI = Mask[l+i];
4243 int BitI1 = Mask[l+i+1];
4244 if (!isUndefOrEqual(BitI, j))
4246 if (!isUndefOrEqual(BitI1, j))
4253 // Match for INSERTI64x4 INSERTF64x4 instructions (src0[0], src1[0]) or
4254 // (src1[0], src0[1]), manipulation with 256-bit sub-vectors
4255 static bool isINSERT64x4Mask(ArrayRef<int> Mask, MVT VT, unsigned int *Imm) {
4256 if (!VT.is512BitVector())
4259 unsigned NumElts = VT.getVectorNumElements();
4260 unsigned HalfSize = NumElts/2;
4261 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, 0)) {
4262 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, NumElts)) {
4267 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, NumElts)) {
4268 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, HalfSize)) {
4276 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
4277 /// specifies a shuffle of elements that is suitable for input to MOVSS,
4278 /// MOVSD, and MOVD, i.e. setting the lowest element.
4279 static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
4280 if (VT.getVectorElementType().getSizeInBits() < 32)
4282 if (!VT.is128BitVector())
4285 unsigned NumElts = VT.getVectorNumElements();
4287 if (!isUndefOrEqual(Mask[0], NumElts))
4290 for (unsigned i = 1; i != NumElts; ++i)
4291 if (!isUndefOrEqual(Mask[i], i))
4297 /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
4298 /// as permutations between 128-bit chunks or halves. As an example: this
4300 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
4301 /// The first half comes from the second half of V1 and the second half from the
4302 /// the second half of V2.
4303 static bool isVPERM2X128Mask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4304 if (!HasFp256 || !VT.is256BitVector())
4307 // The shuffle result is divided into half A and half B. In total the two
4308 // sources have 4 halves, namely: C, D, E, F. The final values of A and
4309 // B must come from C, D, E or F.
4310 unsigned HalfSize = VT.getVectorNumElements()/2;
4311 bool MatchA = false, MatchB = false;
4313 // Check if A comes from one of C, D, E, F.
4314 for (unsigned Half = 0; Half != 4; ++Half) {
4315 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
4321 // Check if B comes from one of C, D, E, F.
4322 for (unsigned Half = 0; Half != 4; ++Half) {
4323 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
4329 return MatchA && MatchB;
4332 /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
4333 /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
4334 static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
4335 MVT VT = SVOp->getSimpleValueType(0);
4337 unsigned HalfSize = VT.getVectorNumElements()/2;
4339 unsigned FstHalf = 0, SndHalf = 0;
4340 for (unsigned i = 0; i < HalfSize; ++i) {
4341 if (SVOp->getMaskElt(i) > 0) {
4342 FstHalf = SVOp->getMaskElt(i)/HalfSize;
4346 for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
4347 if (SVOp->getMaskElt(i) > 0) {
4348 SndHalf = SVOp->getMaskElt(i)/HalfSize;
4353 return (FstHalf | (SndHalf << 4));
4356 // Symetric in-lane mask. Each lane has 4 elements (for imm8)
4357 static bool isPermImmMask(ArrayRef<int> Mask, MVT VT, unsigned& Imm8) {
4358 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4362 unsigned NumElts = VT.getVectorNumElements();
4364 if (VT.is128BitVector() || (VT.is256BitVector() && EltSize == 64)) {
4365 for (unsigned i = 0; i != NumElts; ++i) {
4368 Imm8 |= Mask[i] << (i*2);
4373 unsigned LaneSize = 4;
4374 SmallVector<int, 4> MaskVal(LaneSize, -1);
4376 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4377 for (unsigned i = 0; i != LaneSize; ++i) {
4378 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4382 if (MaskVal[i] < 0) {
4383 MaskVal[i] = Mask[i+l] - l;
4384 Imm8 |= MaskVal[i] << (i*2);
4387 if (Mask[i+l] != (signed)(MaskVal[i]+l))
4394 /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
4395 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
4396 /// Note that VPERMIL mask matching is different depending whether theunderlying
4397 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
4398 /// to the same elements of the low, but to the higher half of the source.
4399 /// In VPERMILPD the two lanes could be shuffled independently of each other
4400 /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
4401 static bool isVPERMILPMask(ArrayRef<int> Mask, MVT VT) {
4402 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4403 if (VT.getSizeInBits() < 256 || EltSize < 32)
4405 bool symetricMaskRequired = (EltSize == 32);
4406 unsigned NumElts = VT.getVectorNumElements();
4408 unsigned NumLanes = VT.getSizeInBits()/128;
4409 unsigned LaneSize = NumElts/NumLanes;
4410 // 2 or 4 elements in one lane
4412 SmallVector<int, 4> ExpectedMaskVal(LaneSize, -1);
4413 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4414 for (unsigned i = 0; i != LaneSize; ++i) {
4415 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4417 if (symetricMaskRequired) {
4418 if (ExpectedMaskVal[i] < 0 && Mask[i+l] >= 0) {
4419 ExpectedMaskVal[i] = Mask[i+l] - l;
4422 if (!isUndefOrEqual(Mask[i+l], ExpectedMaskVal[i]+l))
4430 /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
4431 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
4432 /// element of vector 2 and the other elements to come from vector 1 in order.
4433 static bool isCommutedMOVLMask(ArrayRef<int> Mask, MVT VT,
4434 bool V2IsSplat = false, bool V2IsUndef = false) {
4435 if (!VT.is128BitVector())
4438 unsigned NumOps = VT.getVectorNumElements();
4439 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
4442 if (!isUndefOrEqual(Mask[0], 0))
4445 for (unsigned i = 1; i != NumOps; ++i)
4446 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
4447 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
4448 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
4454 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4455 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
4456 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
4457 static bool isMOVSHDUPMask(ArrayRef<int> Mask, MVT VT,
4458 const X86Subtarget *Subtarget) {
4459 if (!Subtarget->hasSSE3())
4462 unsigned NumElems = VT.getVectorNumElements();
4464 if ((VT.is128BitVector() && NumElems != 4) ||
4465 (VT.is256BitVector() && NumElems != 8) ||
4466 (VT.is512BitVector() && NumElems != 16))
4469 // "i+1" is the value the indexed mask element must have
4470 for (unsigned i = 0; i != NumElems; i += 2)
4471 if (!isUndefOrEqual(Mask[i], i+1) ||
4472 !isUndefOrEqual(Mask[i+1], i+1))
4478 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4479 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
4480 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
4481 static bool isMOVSLDUPMask(ArrayRef<int> Mask, MVT VT,
4482 const X86Subtarget *Subtarget) {
4483 if (!Subtarget->hasSSE3())
4486 unsigned NumElems = VT.getVectorNumElements();
4488 if ((VT.is128BitVector() && NumElems != 4) ||
4489 (VT.is256BitVector() && NumElems != 8) ||
4490 (VT.is512BitVector() && NumElems != 16))
4493 // "i" is the value the indexed mask element must have
4494 for (unsigned i = 0; i != NumElems; i += 2)
4495 if (!isUndefOrEqual(Mask[i], i) ||
4496 !isUndefOrEqual(Mask[i+1], i))
4502 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
4503 /// specifies a shuffle of elements that is suitable for input to 256-bit
4504 /// version of MOVDDUP.
4505 static bool isMOVDDUPYMask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4506 if (!HasFp256 || !VT.is256BitVector())
4509 unsigned NumElts = VT.getVectorNumElements();
4513 for (unsigned i = 0; i != NumElts/2; ++i)
4514 if (!isUndefOrEqual(Mask[i], 0))
4516 for (unsigned i = NumElts/2; i != NumElts; ++i)
4517 if (!isUndefOrEqual(Mask[i], NumElts/2))
4522 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4523 /// specifies a shuffle of elements that is suitable for input to 128-bit
4524 /// version of MOVDDUP.
4525 static bool isMOVDDUPMask(ArrayRef<int> Mask, MVT VT) {
4526 if (!VT.is128BitVector())
4529 unsigned e = VT.getVectorNumElements() / 2;
4530 for (unsigned i = 0; i != e; ++i)
4531 if (!isUndefOrEqual(Mask[i], i))
4533 for (unsigned i = 0; i != e; ++i)
4534 if (!isUndefOrEqual(Mask[e+i], i))
4539 /// isVEXTRACTIndex - Return true if the specified
4540 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
4541 /// suitable for instruction that extract 128 or 256 bit vectors
4542 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4543 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4544 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4547 // The index should be aligned on a vecWidth-bit boundary.
4549 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4551 MVT VT = N->getSimpleValueType(0);
4552 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4553 bool Result = (Index * ElSize) % vecWidth == 0;
4558 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
4559 /// operand specifies a subvector insert that is suitable for input to
4560 /// insertion of 128 or 256-bit subvectors
4561 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4562 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4563 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4565 // The index should be aligned on a vecWidth-bit boundary.
4567 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4569 MVT VT = N->getSimpleValueType(0);
4570 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4571 bool Result = (Index * ElSize) % vecWidth == 0;
4576 bool X86::isVINSERT128Index(SDNode *N) {
4577 return isVINSERTIndex(N, 128);
4580 bool X86::isVINSERT256Index(SDNode *N) {
4581 return isVINSERTIndex(N, 256);
4584 bool X86::isVEXTRACT128Index(SDNode *N) {
4585 return isVEXTRACTIndex(N, 128);
4588 bool X86::isVEXTRACT256Index(SDNode *N) {
4589 return isVEXTRACTIndex(N, 256);
4592 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4593 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4594 /// Handles 128-bit and 256-bit.
4595 static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
4596 MVT VT = N->getSimpleValueType(0);
4598 assert((VT.getSizeInBits() >= 128) &&
4599 "Unsupported vector type for PSHUF/SHUFP");
4601 // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
4602 // independently on 128-bit lanes.
4603 unsigned NumElts = VT.getVectorNumElements();
4604 unsigned NumLanes = VT.getSizeInBits()/128;
4605 unsigned NumLaneElts = NumElts/NumLanes;
4607 assert((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) &&
4608 "Only supports 2, 4 or 8 elements per lane");
4610 unsigned Shift = (NumLaneElts >= 4) ? 1 : 0;
4612 for (unsigned i = 0; i != NumElts; ++i) {
4613 int Elt = N->getMaskElt(i);
4614 if (Elt < 0) continue;
4615 Elt &= NumLaneElts - 1;
4616 unsigned ShAmt = (i << Shift) % 8;
4617 Mask |= Elt << ShAmt;
4623 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4624 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4625 static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
4626 MVT VT = N->getSimpleValueType(0);
4628 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4629 "Unsupported vector type for PSHUFHW");
4631 unsigned NumElts = VT.getVectorNumElements();
4634 for (unsigned l = 0; l != NumElts; l += 8) {
4635 // 8 nodes per lane, but we only care about the last 4.
4636 for (unsigned i = 0; i < 4; ++i) {
4637 int Elt = N->getMaskElt(l+i+4);
4638 if (Elt < 0) continue;
4639 Elt &= 0x3; // only 2-bits.
4640 Mask |= Elt << (i * 2);
4647 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4648 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4649 static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
4650 MVT VT = N->getSimpleValueType(0);
4652 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4653 "Unsupported vector type for PSHUFHW");
4655 unsigned NumElts = VT.getVectorNumElements();
4658 for (unsigned l = 0; l != NumElts; l += 8) {
4659 // 8 nodes per lane, but we only care about the first 4.
4660 for (unsigned i = 0; i < 4; ++i) {
4661 int Elt = N->getMaskElt(l+i);
4662 if (Elt < 0) continue;
4663 Elt &= 0x3; // only 2-bits
4664 Mask |= Elt << (i * 2);
4671 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
4672 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
4673 static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
4674 MVT VT = SVOp->getSimpleValueType(0);
4675 unsigned EltSize = VT.is512BitVector() ? 1 :
4676 VT.getVectorElementType().getSizeInBits() >> 3;
4678 unsigned NumElts = VT.getVectorNumElements();
4679 unsigned NumLanes = VT.is512BitVector() ? 1 : VT.getSizeInBits()/128;
4680 unsigned NumLaneElts = NumElts/NumLanes;
4684 for (i = 0; i != NumElts; ++i) {
4685 Val = SVOp->getMaskElt(i);
4689 if (Val >= (int)NumElts)
4690 Val -= NumElts - NumLaneElts;
4692 assert(Val - i > 0 && "PALIGNR imm should be positive");
4693 return (Val - i) * EltSize;
4696 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4697 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4698 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4699 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4702 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4704 MVT VecVT = N->getOperand(0).getSimpleValueType();
4705 MVT ElVT = VecVT.getVectorElementType();
4707 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4708 return Index / NumElemsPerChunk;
4711 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4712 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4713 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4714 llvm_unreachable("Illegal insert subvector for VINSERT");
4717 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4719 MVT VecVT = N->getSimpleValueType(0);
4720 MVT ElVT = VecVT.getVectorElementType();
4722 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4723 return Index / NumElemsPerChunk;
4726 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4727 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4728 /// and VINSERTI128 instructions.
4729 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4730 return getExtractVEXTRACTImmediate(N, 128);
4733 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
4734 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4735 /// and VINSERTI64x4 instructions.
4736 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4737 return getExtractVEXTRACTImmediate(N, 256);
4740 /// getInsertVINSERT128Immediate - Return the appropriate immediate
4741 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4742 /// and VINSERTI128 instructions.
4743 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4744 return getInsertVINSERTImmediate(N, 128);
4747 /// getInsertVINSERT256Immediate - Return the appropriate immediate
4748 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
4749 /// and VINSERTI64x4 instructions.
4750 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4751 return getInsertVINSERTImmediate(N, 256);
4754 /// isZero - Returns true if Elt is a constant integer zero
4755 static bool isZero(SDValue V) {
4756 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4757 return C && C->isNullValue();
4760 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4762 bool X86::isZeroNode(SDValue Elt) {
4765 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4766 return CFP->getValueAPF().isPosZero();
4770 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
4771 /// match movhlps. The lower half elements should come from upper half of
4772 /// V1 (and in order), and the upper half elements should come from the upper
4773 /// half of V2 (and in order).
4774 static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, MVT VT) {
4775 if (!VT.is128BitVector())
4777 if (VT.getVectorNumElements() != 4)
4779 for (unsigned i = 0, e = 2; i != e; ++i)
4780 if (!isUndefOrEqual(Mask[i], i+2))
4782 for (unsigned i = 2; i != 4; ++i)
4783 if (!isUndefOrEqual(Mask[i], i+4))
4788 /// isScalarLoadToVector - Returns true if the node is a scalar load that
4789 /// is promoted to a vector. It also returns the LoadSDNode by reference if
4791 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = nullptr) {
4792 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
4794 N = N->getOperand(0).getNode();
4795 if (!ISD::isNON_EXTLoad(N))
4798 *LD = cast<LoadSDNode>(N);
4802 // Test whether the given value is a vector value which will be legalized
4804 static bool WillBeConstantPoolLoad(SDNode *N) {
4805 if (N->getOpcode() != ISD::BUILD_VECTOR)
4808 // Check for any non-constant elements.
4809 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
4810 switch (N->getOperand(i).getNode()->getOpcode()) {
4812 case ISD::ConstantFP:
4819 // Vectors of all-zeros and all-ones are materialized with special
4820 // instructions rather than being loaded.
4821 return !ISD::isBuildVectorAllZeros(N) &&
4822 !ISD::isBuildVectorAllOnes(N);
4825 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
4826 /// match movlp{s|d}. The lower half elements should come from lower half of
4827 /// V1 (and in order), and the upper half elements should come from the upper
4828 /// half of V2 (and in order). And since V1 will become the source of the
4829 /// MOVLP, it must be either a vector load or a scalar load to vector.
4830 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
4831 ArrayRef<int> Mask, MVT VT) {
4832 if (!VT.is128BitVector())
4835 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
4837 // Is V2 is a vector load, don't do this transformation. We will try to use
4838 // load folding shufps op.
4839 if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
4842 unsigned NumElems = VT.getVectorNumElements();
4844 if (NumElems != 2 && NumElems != 4)
4846 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4847 if (!isUndefOrEqual(Mask[i], i))
4849 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
4850 if (!isUndefOrEqual(Mask[i], i+NumElems))
4855 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
4856 /// to an zero vector.
4857 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
4858 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
4859 SDValue V1 = N->getOperand(0);
4860 SDValue V2 = N->getOperand(1);
4861 unsigned NumElems = N->getValueType(0).getVectorNumElements();
4862 for (unsigned i = 0; i != NumElems; ++i) {
4863 int Idx = N->getMaskElt(i);
4864 if (Idx >= (int)NumElems) {
4865 unsigned Opc = V2.getOpcode();
4866 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
4868 if (Opc != ISD::BUILD_VECTOR ||
4869 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
4871 } else if (Idx >= 0) {
4872 unsigned Opc = V1.getOpcode();
4873 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
4875 if (Opc != ISD::BUILD_VECTOR ||
4876 !X86::isZeroNode(V1.getOperand(Idx)))
4883 /// getZeroVector - Returns a vector of specified type with all zero elements.
4885 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4886 SelectionDAG &DAG, SDLoc dl) {
4887 assert(VT.isVector() && "Expected a vector type");
4889 // Always build SSE zero vectors as <4 x i32> bitcasted
4890 // to their dest type. This ensures they get CSE'd.
4892 if (VT.is128BitVector()) { // SSE
4893 if (Subtarget->hasSSE2()) { // SSE2
4894 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4895 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4897 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4898 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4900 } else if (VT.is256BitVector()) { // AVX
4901 if (Subtarget->hasInt256()) { // AVX2
4902 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4903 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4904 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4906 // 256-bit logic and arithmetic instructions in AVX are all
4907 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4908 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4909 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4910 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4912 } else if (VT.is512BitVector()) { // AVX-512
4913 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4914 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4915 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4916 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4917 } else if (VT.getScalarType() == MVT::i1) {
4918 assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
4919 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
4920 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
4921 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4923 llvm_unreachable("Unexpected vector type");
4925 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4928 /// getOnesVector - Returns a vector of specified type with all bits set.
4929 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4930 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4931 /// Then bitcast to their original type, ensuring they get CSE'd.
4932 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4934 assert(VT.isVector() && "Expected a vector type");
4936 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
4938 if (VT.is256BitVector()) {
4939 if (HasInt256) { // AVX2
4940 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4941 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4943 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4944 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4946 } else if (VT.is128BitVector()) {
4947 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4949 llvm_unreachable("Unexpected vector type");
4951 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4954 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
4955 /// that point to V2 points to its first element.
4956 static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
4957 for (unsigned i = 0; i != NumElems; ++i) {
4958 if (Mask[i] > (int)NumElems) {
4964 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4965 /// operation of specified width.
4966 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4968 unsigned NumElems = VT.getVectorNumElements();
4969 SmallVector<int, 8> Mask;
4970 Mask.push_back(NumElems);
4971 for (unsigned i = 1; i != NumElems; ++i)
4973 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4976 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4977 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4979 unsigned NumElems = VT.getVectorNumElements();
4980 SmallVector<int, 8> Mask;
4981 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4983 Mask.push_back(i + NumElems);
4985 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4988 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4989 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4991 unsigned NumElems = VT.getVectorNumElements();
4992 SmallVector<int, 8> Mask;
4993 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4994 Mask.push_back(i + Half);
4995 Mask.push_back(i + NumElems + Half);
4997 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5000 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
5001 // a generic shuffle instruction because the target has no such instructions.
5002 // Generate shuffles which repeat i16 and i8 several times until they can be
5003 // represented by v4f32 and then be manipulated by target suported shuffles.
5004 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
5005 MVT VT = V.getSimpleValueType();
5006 int NumElems = VT.getVectorNumElements();
5009 while (NumElems > 4) {
5010 if (EltNo < NumElems/2) {
5011 V = getUnpackl(DAG, dl, VT, V, V);
5013 V = getUnpackh(DAG, dl, VT, V, V);
5014 EltNo -= NumElems/2;
5021 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
5022 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
5023 MVT VT = V.getSimpleValueType();
5026 if (VT.is128BitVector()) {
5027 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
5028 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
5029 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
5031 } else if (VT.is256BitVector()) {
5032 // To use VPERMILPS to splat scalars, the second half of indicies must
5033 // refer to the higher part, which is a duplication of the lower one,
5034 // because VPERMILPS can only handle in-lane permutations.
5035 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
5036 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
5038 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
5039 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
5042 llvm_unreachable("Vector size not supported");
5044 return DAG.getNode(ISD::BITCAST, dl, VT, V);
5047 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
5048 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
5049 MVT SrcVT = SV->getSimpleValueType(0);
5050 SDValue V1 = SV->getOperand(0);
5053 int EltNo = SV->getSplatIndex();
5054 int NumElems = SrcVT.getVectorNumElements();
5055 bool Is256BitVec = SrcVT.is256BitVector();
5057 assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&
5058 "Unknown how to promote splat for type");
5060 // Extract the 128-bit part containing the splat element and update
5061 // the splat element index when it refers to the higher register.
5063 V1 = Extract128BitVector(V1, EltNo, DAG, dl);
5064 if (EltNo >= NumElems/2)
5065 EltNo -= NumElems/2;
5068 // All i16 and i8 vector types can't be used directly by a generic shuffle
5069 // instruction because the target has no such instruction. Generate shuffles
5070 // which repeat i16 and i8 several times until they fit in i32, and then can
5071 // be manipulated by target suported shuffles.
5072 MVT EltVT = SrcVT.getVectorElementType();
5073 if (EltVT == MVT::i8 || EltVT == MVT::i16)
5074 V1 = PromoteSplati8i16(V1, DAG, EltNo);
5076 // Recreate the 256-bit vector and place the same 128-bit vector
5077 // into the low and high part. This is necessary because we want
5078 // to use VPERM* to shuffle the vectors
5080 V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
5083 return getLegalSplat(DAG, V1, EltNo);
5086 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
5087 /// vector of zero or undef vector. This produces a shuffle where the low
5088 /// element of V2 is swizzled into the zero/undef vector, landing at element
5089 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
5090 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
5092 const X86Subtarget *Subtarget,
5093 SelectionDAG &DAG) {
5094 MVT VT = V2.getSimpleValueType();
5096 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
5097 unsigned NumElems = VT.getVectorNumElements();
5098 SmallVector<int, 16> MaskVec;
5099 for (unsigned i = 0; i != NumElems; ++i)
5100 // If this is the insertion idx, put the low elt of V2 here.
5101 MaskVec.push_back(i == Idx ? NumElems : i);
5102 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
5105 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
5106 /// target specific opcode. Returns true if the Mask could be calculated.
5107 /// Sets IsUnary to true if only uses one source.
5108 static bool getTargetShuffleMask(SDNode *N, MVT VT,
5109 SmallVectorImpl<int> &Mask, bool &IsUnary) {
5110 unsigned NumElems = VT.getVectorNumElements();
5114 switch(N->getOpcode()) {
5116 ImmN = N->getOperand(N->getNumOperands()-1);
5117 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5119 case X86ISD::UNPCKH:
5120 DecodeUNPCKHMask(VT, Mask);
5122 case X86ISD::UNPCKL:
5123 DecodeUNPCKLMask(VT, Mask);
5125 case X86ISD::MOVHLPS:
5126 DecodeMOVHLPSMask(NumElems, Mask);
5128 case X86ISD::MOVLHPS:
5129 DecodeMOVLHPSMask(NumElems, Mask);
5131 case X86ISD::PALIGNR:
5132 ImmN = N->getOperand(N->getNumOperands()-1);
5133 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5135 case X86ISD::PSHUFD:
5136 case X86ISD::VPERMILP:
5137 ImmN = N->getOperand(N->getNumOperands()-1);
5138 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5141 case X86ISD::PSHUFHW:
5142 ImmN = N->getOperand(N->getNumOperands()-1);
5143 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5146 case X86ISD::PSHUFLW:
5147 ImmN = N->getOperand(N->getNumOperands()-1);
5148 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5151 case X86ISD::VPERMI:
5152 ImmN = N->getOperand(N->getNumOperands()-1);
5153 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5157 case X86ISD::MOVSD: {
5158 // The index 0 always comes from the first element of the second source,
5159 // this is why MOVSS and MOVSD are used in the first place. The other
5160 // elements come from the other positions of the first source vector
5161 Mask.push_back(NumElems);
5162 for (unsigned i = 1; i != NumElems; ++i) {
5167 case X86ISD::VPERM2X128:
5168 ImmN = N->getOperand(N->getNumOperands()-1);
5169 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5170 if (Mask.empty()) return false;
5172 case X86ISD::MOVDDUP:
5173 case X86ISD::MOVLHPD:
5174 case X86ISD::MOVLPD:
5175 case X86ISD::MOVLPS:
5176 case X86ISD::MOVSHDUP:
5177 case X86ISD::MOVSLDUP:
5178 // Not yet implemented
5180 default: llvm_unreachable("unknown target shuffle node");
5186 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5187 /// element of the result of the vector shuffle.
5188 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
5191 return SDValue(); // Limit search depth.
5193 SDValue V = SDValue(N, 0);
5194 EVT VT = V.getValueType();
5195 unsigned Opcode = V.getOpcode();
5197 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
5198 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
5199 int Elt = SV->getMaskElt(Index);
5202 return DAG.getUNDEF(VT.getVectorElementType());
5204 unsigned NumElems = VT.getVectorNumElements();
5205 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
5206 : SV->getOperand(1);
5207 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
5210 // Recurse into target specific vector shuffles to find scalars.
5211 if (isTargetShuffle(Opcode)) {
5212 MVT ShufVT = V.getSimpleValueType();
5213 unsigned NumElems = ShufVT.getVectorNumElements();
5214 SmallVector<int, 16> ShuffleMask;
5217 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
5220 int Elt = ShuffleMask[Index];
5222 return DAG.getUNDEF(ShufVT.getVectorElementType());
5224 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
5226 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5230 // Actual nodes that may contain scalar elements
5231 if (Opcode == ISD::BITCAST) {
5232 V = V.getOperand(0);
5233 EVT SrcVT = V.getValueType();
5234 unsigned NumElems = VT.getVectorNumElements();
5236 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5240 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5241 return (Index == 0) ? V.getOperand(0)
5242 : DAG.getUNDEF(VT.getVectorElementType());
5244 if (V.getOpcode() == ISD::BUILD_VECTOR)
5245 return V.getOperand(Index);
5250 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
5251 /// shuffle operation which come from a consecutively from a zero. The
5252 /// search can start in two different directions, from left or right.
5253 /// We count undefs as zeros until PreferredNum is reached.
5254 static unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp,
5255 unsigned NumElems, bool ZerosFromLeft,
5257 unsigned PreferredNum = -1U) {
5258 unsigned NumZeros = 0;
5259 for (unsigned i = 0; i != NumElems; ++i) {
5260 unsigned Index = ZerosFromLeft ? i : NumElems - i - 1;
5261 SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
5265 if (X86::isZeroNode(Elt))
5267 else if (Elt.getOpcode() == ISD::UNDEF) // Undef as zero up to PreferredNum.
5268 NumZeros = std::min(NumZeros + 1, PreferredNum);
5276 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
5277 /// correspond consecutively to elements from one of the vector operands,
5278 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
5280 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
5281 unsigned MaskI, unsigned MaskE, unsigned OpIdx,
5282 unsigned NumElems, unsigned &OpNum) {
5283 bool SeenV1 = false;
5284 bool SeenV2 = false;
5286 for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
5287 int Idx = SVOp->getMaskElt(i);
5288 // Ignore undef indicies
5292 if (Idx < (int)NumElems)
5297 // Only accept consecutive elements from the same vector
5298 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
5302 OpNum = SeenV1 ? 0 : 1;
5306 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
5307 /// logical left shift of a vector.
5308 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5309 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5311 SVOp->getSimpleValueType(0).getVectorNumElements();
5312 unsigned NumZeros = getNumOfConsecutiveZeros(
5313 SVOp, NumElems, false /* check zeros from right */, DAG,
5314 SVOp->getMaskElt(0));
5320 // Considering the elements in the mask that are not consecutive zeros,
5321 // check if they consecutively come from only one of the source vectors.
5323 // V1 = {X, A, B, C} 0
5325 // vector_shuffle V1, V2 <1, 2, 3, X>
5327 if (!isShuffleMaskConsecutive(SVOp,
5328 0, // Mask Start Index
5329 NumElems-NumZeros, // Mask End Index(exclusive)
5330 NumZeros, // Where to start looking in the src vector
5331 NumElems, // Number of elements in vector
5332 OpSrc)) // Which source operand ?
5337 ShVal = SVOp->getOperand(OpSrc);
5341 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
5342 /// logical left shift of a vector.
5343 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5344 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5346 SVOp->getSimpleValueType(0).getVectorNumElements();
5347 unsigned NumZeros = getNumOfConsecutiveZeros(
5348 SVOp, NumElems, true /* check zeros from left */, DAG,
5349 NumElems - SVOp->getMaskElt(NumElems - 1) - 1);
5355 // Considering the elements in the mask that are not consecutive zeros,
5356 // check if they consecutively come from only one of the source vectors.
5358 // 0 { A, B, X, X } = V2
5360 // vector_shuffle V1, V2 <X, X, 4, 5>
5362 if (!isShuffleMaskConsecutive(SVOp,
5363 NumZeros, // Mask Start Index
5364 NumElems, // Mask End Index(exclusive)
5365 0, // Where to start looking in the src vector
5366 NumElems, // Number of elements in vector
5367 OpSrc)) // Which source operand ?
5372 ShVal = SVOp->getOperand(OpSrc);
5376 /// isVectorShift - Returns true if the shuffle can be implemented as a
5377 /// logical left or right shift of a vector.
5378 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5379 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5380 // Although the logic below support any bitwidth size, there are no
5381 // shift instructions which handle more than 128-bit vectors.
5382 if (!SVOp->getSimpleValueType(0).is128BitVector())
5385 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
5386 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
5392 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
5394 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5395 unsigned NumNonZero, unsigned NumZero,
5397 const X86Subtarget* Subtarget,
5398 const TargetLowering &TLI) {
5405 for (unsigned i = 0; i < 16; ++i) {
5406 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5407 if (ThisIsNonZero && First) {
5409 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5411 V = DAG.getUNDEF(MVT::v8i16);
5416 SDValue ThisElt, LastElt;
5417 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5418 if (LastIsNonZero) {
5419 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5420 MVT::i16, Op.getOperand(i-1));
5422 if (ThisIsNonZero) {
5423 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5424 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5425 ThisElt, DAG.getConstant(8, MVT::i8));
5427 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5431 if (ThisElt.getNode())
5432 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5433 DAG.getIntPtrConstant(i/2));
5437 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
5440 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
5442 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5443 unsigned NumNonZero, unsigned NumZero,
5445 const X86Subtarget* Subtarget,
5446 const TargetLowering &TLI) {
5453 for (unsigned i = 0; i < 8; ++i) {
5454 bool isNonZero = (NonZeros & (1 << i)) != 0;
5458 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5460 V = DAG.getUNDEF(MVT::v8i16);
5463 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5464 MVT::v8i16, V, Op.getOperand(i),
5465 DAG.getIntPtrConstant(i));
5472 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
5473 static SDValue LowerBuildVectorv4x32(SDValue Op, unsigned NumElems,
5474 unsigned NonZeros, unsigned NumNonZero,
5475 unsigned NumZero, SelectionDAG &DAG,
5476 const X86Subtarget *Subtarget,
5477 const TargetLowering &TLI) {
5478 // We know there's at least one non-zero element
5479 unsigned FirstNonZeroIdx = 0;
5480 SDValue FirstNonZero = Op->getOperand(FirstNonZeroIdx);
5481 while (FirstNonZero.getOpcode() == ISD::UNDEF ||
5482 X86::isZeroNode(FirstNonZero)) {
5484 FirstNonZero = Op->getOperand(FirstNonZeroIdx);
5487 if (FirstNonZero.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5488 !isa<ConstantSDNode>(FirstNonZero.getOperand(1)))
5491 SDValue V = FirstNonZero.getOperand(0);
5492 MVT VVT = V.getSimpleValueType();
5493 if (!Subtarget->hasSSE41() || (VVT != MVT::v4f32 && VVT != MVT::v4i32))
5496 unsigned FirstNonZeroDst =
5497 cast<ConstantSDNode>(FirstNonZero.getOperand(1))->getZExtValue();
5498 unsigned CorrectIdx = FirstNonZeroDst == FirstNonZeroIdx;
5499 unsigned IncorrectIdx = CorrectIdx ? -1U : FirstNonZeroIdx;
5500 unsigned IncorrectDst = CorrectIdx ? -1U : FirstNonZeroDst;
5502 for (unsigned Idx = FirstNonZeroIdx + 1; Idx < NumElems; ++Idx) {
5503 SDValue Elem = Op.getOperand(Idx);
5504 if (Elem.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elem))
5507 // TODO: What else can be here? Deal with it.
5508 if (Elem.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5511 // TODO: Some optimizations are still possible here
5512 // ex: Getting one element from a vector, and the rest from another.
5513 if (Elem.getOperand(0) != V)
5516 unsigned Dst = cast<ConstantSDNode>(Elem.getOperand(1))->getZExtValue();
5519 else if (IncorrectIdx == -1U) {
5523 // There was already one element with an incorrect index.
5524 // We can't optimize this case to an insertps.
5528 if (NumNonZero == CorrectIdx || NumNonZero == CorrectIdx + 1) {
5530 EVT VT = Op.getSimpleValueType();
5531 unsigned ElementMoveMask = 0;
5532 if (IncorrectIdx == -1U)
5533 ElementMoveMask = FirstNonZeroIdx << 6 | FirstNonZeroIdx << 4;
5535 ElementMoveMask = IncorrectDst << 6 | IncorrectIdx << 4;
5537 SDValue InsertpsMask =
5538 DAG.getIntPtrConstant(ElementMoveMask | (~NonZeros & 0xf));
5539 return DAG.getNode(X86ISD::INSERTPS, dl, VT, V, V, InsertpsMask);
5545 /// getVShift - Return a vector logical shift node.
5547 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5548 unsigned NumBits, SelectionDAG &DAG,
5549 const TargetLowering &TLI, SDLoc dl) {
5550 assert(VT.is128BitVector() && "Unknown type for VShift");
5551 EVT ShVT = MVT::v2i64;
5552 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5553 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
5554 return DAG.getNode(ISD::BITCAST, dl, VT,
5555 DAG.getNode(Opc, dl, ShVT, SrcOp,
5556 DAG.getConstant(NumBits,
5557 TLI.getScalarShiftAmountTy(SrcOp.getValueType()))));
5561 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5563 // Check if the scalar load can be widened into a vector load. And if
5564 // the address is "base + cst" see if the cst can be "absorbed" into
5565 // the shuffle mask.
5566 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5567 SDValue Ptr = LD->getBasePtr();
5568 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5570 EVT PVT = LD->getValueType(0);
5571 if (PVT != MVT::i32 && PVT != MVT::f32)
5576 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5577 FI = FINode->getIndex();
5579 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5580 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5581 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5582 Offset = Ptr.getConstantOperandVal(1);
5583 Ptr = Ptr.getOperand(0);
5588 // FIXME: 256-bit vector instructions don't require a strict alignment,
5589 // improve this code to support it better.
5590 unsigned RequiredAlign = VT.getSizeInBits()/8;
5591 SDValue Chain = LD->getChain();
5592 // Make sure the stack object alignment is at least 16 or 32.
5593 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5594 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5595 if (MFI->isFixedObjectIndex(FI)) {
5596 // Can't change the alignment. FIXME: It's possible to compute
5597 // the exact stack offset and reference FI + adjust offset instead.
5598 // If someone *really* cares about this. That's the way to implement it.
5601 MFI->setObjectAlignment(FI, RequiredAlign);
5605 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5606 // Ptr + (Offset & ~15).
5609 if ((Offset % RequiredAlign) & 3)
5611 int64_t StartOffset = Offset & ~(RequiredAlign-1);
5613 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
5614 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
5616 int EltNo = (Offset - StartOffset) >> 2;
5617 unsigned NumElems = VT.getVectorNumElements();
5619 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5620 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5621 LD->getPointerInfo().getWithOffset(StartOffset),
5622 false, false, false, 0);
5624 SmallVector<int, 8> Mask;
5625 for (unsigned i = 0; i != NumElems; ++i)
5626 Mask.push_back(EltNo);
5628 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5634 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
5635 /// vector of type 'VT', see if the elements can be replaced by a single large
5636 /// load which has the same value as a build_vector whose operands are 'elts'.
5638 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5640 /// FIXME: we'd also like to handle the case where the last elements are zero
5641 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5642 /// There's even a handy isZeroNode for that purpose.
5643 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
5644 SDLoc &DL, SelectionDAG &DAG,
5645 bool isAfterLegalize) {
5646 EVT EltVT = VT.getVectorElementType();
5647 unsigned NumElems = Elts.size();
5649 LoadSDNode *LDBase = nullptr;
5650 unsigned LastLoadedElt = -1U;
5652 // For each element in the initializer, see if we've found a load or an undef.
5653 // If we don't find an initial load element, or later load elements are
5654 // non-consecutive, bail out.
5655 for (unsigned i = 0; i < NumElems; ++i) {
5656 SDValue Elt = Elts[i];
5658 if (!Elt.getNode() ||
5659 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5662 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5664 LDBase = cast<LoadSDNode>(Elt.getNode());
5668 if (Elt.getOpcode() == ISD::UNDEF)
5671 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5672 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
5677 // If we have found an entire vector of loads and undefs, then return a large
5678 // load of the entire vector width starting at the base pointer. If we found
5679 // consecutive loads for the low half, generate a vzext_load node.
5680 if (LastLoadedElt == NumElems - 1) {
5682 if (isAfterLegalize &&
5683 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5686 SDValue NewLd = SDValue();
5688 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
5689 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5690 LDBase->getPointerInfo(),
5691 LDBase->isVolatile(), LDBase->isNonTemporal(),
5692 LDBase->isInvariant(), 0);
5693 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5694 LDBase->getPointerInfo(),
5695 LDBase->isVolatile(), LDBase->isNonTemporal(),
5696 LDBase->isInvariant(), LDBase->getAlignment());
5698 if (LDBase->hasAnyUseOfValue(1)) {
5699 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5701 SDValue(NewLd.getNode(), 1));
5702 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5703 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5704 SDValue(NewLd.getNode(), 1));
5709 if (NumElems == 4 && LastLoadedElt == 1 &&
5710 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5711 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5712 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5714 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5715 LDBase->getPointerInfo(),
5716 LDBase->getAlignment(),
5717 false/*isVolatile*/, true/*ReadMem*/,
5720 // Make sure the newly-created LOAD is in the same position as LDBase in
5721 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5722 // update uses of LDBase's output chain to use the TokenFactor.
5723 if (LDBase->hasAnyUseOfValue(1)) {
5724 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5725 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5726 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5727 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5728 SDValue(ResNode.getNode(), 1));
5731 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
5736 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5737 /// to generate a splat value for the following cases:
5738 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5739 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5740 /// a scalar load, or a constant.
5741 /// The VBROADCAST node is returned when a pattern is found,
5742 /// or SDValue() otherwise.
5743 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5744 SelectionDAG &DAG) {
5745 if (!Subtarget->hasFp256())
5748 MVT VT = Op.getSimpleValueType();
5751 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5752 "Unsupported vector type for broadcast.");
5757 switch (Op.getOpcode()) {
5759 // Unknown pattern found.
5762 case ISD::BUILD_VECTOR: {
5763 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5764 BitVector UndefElements;
5765 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5767 // We need a splat of a single value to use broadcast, and it doesn't
5768 // make any sense if the value is only in one element of the vector.
5769 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5773 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5774 Ld.getOpcode() == ISD::ConstantFP);
5776 // Make sure that all of the users of a non-constant load are from the
5777 // BUILD_VECTOR node.
5778 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5783 case ISD::VECTOR_SHUFFLE: {
5784 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5786 // Shuffles must have a splat mask where the first element is
5788 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5791 SDValue Sc = Op.getOperand(0);
5792 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5793 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5795 if (!Subtarget->hasInt256())
5798 // Use the register form of the broadcast instruction available on AVX2.
5799 if (VT.getSizeInBits() >= 256)
5800 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5801 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5804 Ld = Sc.getOperand(0);
5805 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5806 Ld.getOpcode() == ISD::ConstantFP);
5808 // The scalar_to_vector node and the suspected
5809 // load node must have exactly one user.
5810 // Constants may have multiple users.
5812 // AVX-512 has register version of the broadcast
5813 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5814 Ld.getValueType().getSizeInBits() >= 32;
5815 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5822 bool IsGE256 = (VT.getSizeInBits() >= 256);
5824 // Handle the broadcasting a single constant scalar from the constant pool
5825 // into a vector. On Sandybridge it is still better to load a constant vector
5826 // from the constant pool and not to broadcast it from a scalar.
5827 if (ConstSplatVal && Subtarget->hasInt256()) {
5828 EVT CVT = Ld.getValueType();
5829 assert(!CVT.isVector() && "Must not broadcast a vector type");
5830 unsigned ScalarSize = CVT.getSizeInBits();
5832 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)) {
5833 const Constant *C = nullptr;
5834 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5835 C = CI->getConstantIntValue();
5836 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5837 C = CF->getConstantFPValue();
5839 assert(C && "Invalid constant type");
5841 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5842 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
5843 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5844 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5845 MachinePointerInfo::getConstantPool(),
5846 false, false, false, Alignment);
5848 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5852 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5853 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5855 // Handle AVX2 in-register broadcasts.
5856 if (!IsLoad && Subtarget->hasInt256() &&
5857 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5858 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5860 // The scalar source must be a normal load.
5864 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64))
5865 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5867 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5868 // double since there is no vbroadcastsd xmm
5869 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5870 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5871 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5874 // Unsupported broadcast.
5878 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5879 /// underlying vector and index.
5881 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5883 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5885 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5886 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5889 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5891 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5893 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5894 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5897 // In this case the vector is the extract_subvector expression and the index
5898 // is 2, as specified by the shuffle.
5899 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5900 SDValue ShuffleVec = SVOp->getOperand(0);
5901 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5902 assert(ShuffleVecVT.getVectorElementType() ==
5903 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5905 int ShuffleIdx = SVOp->getMaskElt(Idx);
5906 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5907 ExtractedFromVec = ShuffleVec;
5913 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5914 MVT VT = Op.getSimpleValueType();
5916 // Skip if insert_vec_elt is not supported.
5917 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5918 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5922 unsigned NumElems = Op.getNumOperands();
5926 SmallVector<unsigned, 4> InsertIndices;
5927 SmallVector<int, 8> Mask(NumElems, -1);
5929 for (unsigned i = 0; i != NumElems; ++i) {
5930 unsigned Opc = Op.getOperand(i).getOpcode();
5932 if (Opc == ISD::UNDEF)
5935 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5936 // Quit if more than 1 elements need inserting.
5937 if (InsertIndices.size() > 1)
5940 InsertIndices.push_back(i);
5944 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5945 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5946 // Quit if non-constant index.
5947 if (!isa<ConstantSDNode>(ExtIdx))
5949 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5951 // Quit if extracted from vector of different type.
5952 if (ExtractedFromVec.getValueType() != VT)
5955 if (!VecIn1.getNode())
5956 VecIn1 = ExtractedFromVec;
5957 else if (VecIn1 != ExtractedFromVec) {
5958 if (!VecIn2.getNode())
5959 VecIn2 = ExtractedFromVec;
5960 else if (VecIn2 != ExtractedFromVec)
5961 // Quit if more than 2 vectors to shuffle
5965 if (ExtractedFromVec == VecIn1)
5967 else if (ExtractedFromVec == VecIn2)
5968 Mask[i] = Idx + NumElems;
5971 if (!VecIn1.getNode())
5974 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5975 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5976 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5977 unsigned Idx = InsertIndices[i];
5978 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5979 DAG.getIntPtrConstant(Idx));
5985 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5987 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5989 MVT VT = Op.getSimpleValueType();
5990 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
5991 "Unexpected type in LowerBUILD_VECTORvXi1!");
5994 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5995 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
5996 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5997 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6000 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
6001 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
6002 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
6003 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6006 bool AllContants = true;
6007 uint64_t Immediate = 0;
6008 int NonConstIdx = -1;
6009 bool IsSplat = true;
6010 unsigned NumNonConsts = 0;
6011 unsigned NumConsts = 0;
6012 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
6013 SDValue In = Op.getOperand(idx);
6014 if (In.getOpcode() == ISD::UNDEF)
6016 if (!isa<ConstantSDNode>(In)) {
6017 AllContants = false;
6023 if (cast<ConstantSDNode>(In)->getZExtValue())
6024 Immediate |= (1ULL << idx);
6026 if (In != Op.getOperand(0))
6031 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
6032 DAG.getConstant(Immediate, MVT::i16));
6033 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
6034 DAG.getIntPtrConstant(0));
6037 if (NumNonConsts == 1 && NonConstIdx != 0) {
6040 SDValue VecAsImm = DAG.getConstant(Immediate,
6041 MVT::getIntegerVT(VT.getSizeInBits()));
6042 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
6045 DstVec = DAG.getUNDEF(VT);
6046 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
6047 Op.getOperand(NonConstIdx),
6048 DAG.getIntPtrConstant(NonConstIdx));
6050 if (!IsSplat && (NonConstIdx != 0))
6051 llvm_unreachable("Unsupported BUILD_VECTOR operation");
6052 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
6055 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6056 DAG.getConstant(-1, SelectVT),
6057 DAG.getConstant(0, SelectVT));
6059 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6060 DAG.getConstant((Immediate | 1), SelectVT),
6061 DAG.getConstant(Immediate, SelectVT));
6062 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
6065 /// \brief Return true if \p N implements a horizontal binop and return the
6066 /// operands for the horizontal binop into V0 and V1.
6068 /// This is a helper function of PerformBUILD_VECTORCombine.
6069 /// This function checks that the build_vector \p N in input implements a
6070 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
6071 /// operation to match.
6072 /// For example, if \p Opcode is equal to ISD::ADD, then this function
6073 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
6074 /// is equal to ISD::SUB, then this function checks if this is a horizontal
6077 /// This function only analyzes elements of \p N whose indices are
6078 /// in range [BaseIdx, LastIdx).
6079 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
6081 unsigned BaseIdx, unsigned LastIdx,
6082 SDValue &V0, SDValue &V1) {
6083 EVT VT = N->getValueType(0);
6085 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
6086 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
6087 "Invalid Vector in input!");
6089 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
6090 bool CanFold = true;
6091 unsigned ExpectedVExtractIdx = BaseIdx;
6092 unsigned NumElts = LastIdx - BaseIdx;
6093 V0 = DAG.getUNDEF(VT);
6094 V1 = DAG.getUNDEF(VT);
6096 // Check if N implements a horizontal binop.
6097 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
6098 SDValue Op = N->getOperand(i + BaseIdx);
6101 if (Op->getOpcode() == ISD::UNDEF) {
6102 // Update the expected vector extract index.
6103 if (i * 2 == NumElts)
6104 ExpectedVExtractIdx = BaseIdx;
6105 ExpectedVExtractIdx += 2;
6109 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
6114 SDValue Op0 = Op.getOperand(0);
6115 SDValue Op1 = Op.getOperand(1);
6117 // Try to match the following pattern:
6118 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
6119 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6120 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6121 Op0.getOperand(0) == Op1.getOperand(0) &&
6122 isa<ConstantSDNode>(Op0.getOperand(1)) &&
6123 isa<ConstantSDNode>(Op1.getOperand(1)));
6127 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6128 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
6130 if (i * 2 < NumElts) {
6131 if (V0.getOpcode() == ISD::UNDEF)
6132 V0 = Op0.getOperand(0);
6134 if (V1.getOpcode() == ISD::UNDEF)
6135 V1 = Op0.getOperand(0);
6136 if (i * 2 == NumElts)
6137 ExpectedVExtractIdx = BaseIdx;
6140 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
6141 if (I0 == ExpectedVExtractIdx)
6142 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
6143 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
6144 // Try to match the following dag sequence:
6145 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
6146 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
6150 ExpectedVExtractIdx += 2;
6156 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
6157 /// a concat_vector.
6159 /// This is a helper function of PerformBUILD_VECTORCombine.
6160 /// This function expects two 256-bit vectors called V0 and V1.
6161 /// At first, each vector is split into two separate 128-bit vectors.
6162 /// Then, the resulting 128-bit vectors are used to implement two
6163 /// horizontal binary operations.
6165 /// The kind of horizontal binary operation is defined by \p X86Opcode.
6167 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
6168 /// the two new horizontal binop.
6169 /// When Mode is set, the first horizontal binop dag node would take as input
6170 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
6171 /// horizontal binop dag node would take as input the lower 128-bit of V1
6172 /// and the upper 128-bit of V1.
6174 /// HADD V0_LO, V0_HI
6175 /// HADD V1_LO, V1_HI
6177 /// Otherwise, the first horizontal binop dag node takes as input the lower
6178 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
6179 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
6181 /// HADD V0_LO, V1_LO
6182 /// HADD V0_HI, V1_HI
6184 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
6185 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
6186 /// the upper 128-bits of the result.
6187 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
6188 SDLoc DL, SelectionDAG &DAG,
6189 unsigned X86Opcode, bool Mode,
6190 bool isUndefLO, bool isUndefHI) {
6191 EVT VT = V0.getValueType();
6192 assert(VT.is256BitVector() && VT == V1.getValueType() &&
6193 "Invalid nodes in input!");
6195 unsigned NumElts = VT.getVectorNumElements();
6196 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
6197 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
6198 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
6199 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
6200 EVT NewVT = V0_LO.getValueType();
6202 SDValue LO = DAG.getUNDEF(NewVT);
6203 SDValue HI = DAG.getUNDEF(NewVT);
6206 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6207 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
6208 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
6209 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
6210 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
6212 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6213 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
6214 V1_LO->getOpcode() != ISD::UNDEF))
6215 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
6217 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
6218 V1_HI->getOpcode() != ISD::UNDEF))
6219 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
6222 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
6225 /// \brief Try to fold a build_vector that performs an 'addsub' into the
6226 /// sequence of 'vadd + vsub + blendi'.
6227 static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
6228 const X86Subtarget *Subtarget) {
6230 EVT VT = BV->getValueType(0);
6231 unsigned NumElts = VT.getVectorNumElements();
6232 SDValue InVec0 = DAG.getUNDEF(VT);
6233 SDValue InVec1 = DAG.getUNDEF(VT);
6235 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
6236 VT == MVT::v2f64) && "build_vector with an invalid type found!");
6238 // Don't try to emit a VSELECT that cannot be lowered into a blend.
6239 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6240 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
6243 // Odd-numbered elements in the input build vector are obtained from
6244 // adding two integer/float elements.
6245 // Even-numbered elements in the input build vector are obtained from
6246 // subtracting two integer/float elements.
6247 unsigned ExpectedOpcode = ISD::FSUB;
6248 unsigned NextExpectedOpcode = ISD::FADD;
6249 bool AddFound = false;
6250 bool SubFound = false;
6252 for (unsigned i = 0, e = NumElts; i != e; i++) {
6253 SDValue Op = BV->getOperand(i);
6255 // Skip 'undef' values.
6256 unsigned Opcode = Op.getOpcode();
6257 if (Opcode == ISD::UNDEF) {
6258 std::swap(ExpectedOpcode, NextExpectedOpcode);
6262 // Early exit if we found an unexpected opcode.
6263 if (Opcode != ExpectedOpcode)
6266 SDValue Op0 = Op.getOperand(0);
6267 SDValue Op1 = Op.getOperand(1);
6269 // Try to match the following pattern:
6270 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6271 // Early exit if we cannot match that sequence.
6272 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6273 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6274 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6275 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6276 Op0.getOperand(1) != Op1.getOperand(1))
6279 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6283 // We found a valid add/sub node. Update the information accordingly.
6289 // Update InVec0 and InVec1.
6290 if (InVec0.getOpcode() == ISD::UNDEF)
6291 InVec0 = Op0.getOperand(0);
6292 if (InVec1.getOpcode() == ISD::UNDEF)
6293 InVec1 = Op1.getOperand(0);
6295 // Make sure that operands in input to each add/sub node always
6296 // come from a same pair of vectors.
6297 if (InVec0 != Op0.getOperand(0)) {
6298 if (ExpectedOpcode == ISD::FSUB)
6301 // FADD is commutable. Try to commute the operands
6302 // and then test again.
6303 std::swap(Op0, Op1);
6304 if (InVec0 != Op0.getOperand(0))
6308 if (InVec1 != Op1.getOperand(0))
6311 // Update the pair of expected opcodes.
6312 std::swap(ExpectedOpcode, NextExpectedOpcode);
6315 // Don't try to fold this build_vector into a VSELECT if it has
6316 // too many UNDEF operands.
6317 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6318 InVec1.getOpcode() != ISD::UNDEF) {
6319 // Emit a sequence of vector add and sub followed by a VSELECT.
6320 // The new VSELECT will be lowered into a BLENDI.
6321 // At ISel stage, we pattern-match the sequence 'add + sub + BLENDI'
6322 // and emit a single ADDSUB instruction.
6323 SDValue Sub = DAG.getNode(ExpectedOpcode, DL, VT, InVec0, InVec1);
6324 SDValue Add = DAG.getNode(NextExpectedOpcode, DL, VT, InVec0, InVec1);
6326 // Construct the VSELECT mask.
6327 EVT MaskVT = VT.changeVectorElementTypeToInteger();
6328 EVT SVT = MaskVT.getVectorElementType();
6329 unsigned SVTBits = SVT.getSizeInBits();
6330 SmallVector<SDValue, 8> Ops;
6332 for (unsigned i = 0, e = NumElts; i != e; ++i) {
6333 APInt Value = i & 1 ? APInt::getNullValue(SVTBits) :
6334 APInt::getAllOnesValue(SVTBits);
6335 SDValue Constant = DAG.getConstant(Value, SVT);
6336 Ops.push_back(Constant);
6339 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVT, Ops);
6340 return DAG.getSelect(DL, VT, Mask, Sub, Add);
6346 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
6347 const X86Subtarget *Subtarget) {
6349 EVT VT = N->getValueType(0);
6350 unsigned NumElts = VT.getVectorNumElements();
6351 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
6352 SDValue InVec0, InVec1;
6354 // Try to match an ADDSUB.
6355 if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
6356 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
6357 SDValue Value = matchAddSub(BV, DAG, Subtarget);
6358 if (Value.getNode())
6362 // Try to match horizontal ADD/SUB.
6363 unsigned NumUndefsLO = 0;
6364 unsigned NumUndefsHI = 0;
6365 unsigned Half = NumElts/2;
6367 // Count the number of UNDEF operands in the build_vector in input.
6368 for (unsigned i = 0, e = Half; i != e; ++i)
6369 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6372 for (unsigned i = Half, e = NumElts; i != e; ++i)
6373 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6376 // Early exit if this is either a build_vector of all UNDEFs or all the
6377 // operands but one are UNDEF.
6378 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6381 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6382 // Try to match an SSE3 float HADD/HSUB.
6383 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6384 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6386 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6387 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6388 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6389 // Try to match an SSSE3 integer HADD/HSUB.
6390 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6391 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6393 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6394 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6397 if (!Subtarget->hasAVX())
6400 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6401 // Try to match an AVX horizontal add/sub of packed single/double
6402 // precision floating point values from 256-bit vectors.
6403 SDValue InVec2, InVec3;
6404 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6405 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6406 ((InVec0.getOpcode() == ISD::UNDEF ||
6407 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6408 ((InVec1.getOpcode() == ISD::UNDEF ||
6409 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6410 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6412 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6413 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6414 ((InVec0.getOpcode() == ISD::UNDEF ||
6415 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6416 ((InVec1.getOpcode() == ISD::UNDEF ||
6417 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6418 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6419 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6420 // Try to match an AVX2 horizontal add/sub of signed integers.
6421 SDValue InVec2, InVec3;
6423 bool CanFold = true;
6425 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6426 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6427 ((InVec0.getOpcode() == ISD::UNDEF ||
6428 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6429 ((InVec1.getOpcode() == ISD::UNDEF ||
6430 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6431 X86Opcode = X86ISD::HADD;
6432 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6433 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6434 ((InVec0.getOpcode() == ISD::UNDEF ||
6435 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6436 ((InVec1.getOpcode() == ISD::UNDEF ||
6437 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6438 X86Opcode = X86ISD::HSUB;
6443 // Fold this build_vector into a single horizontal add/sub.
6444 // Do this only if the target has AVX2.
6445 if (Subtarget->hasAVX2())
6446 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6448 // Do not try to expand this build_vector into a pair of horizontal
6449 // add/sub if we can emit a pair of scalar add/sub.
6450 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6453 // Convert this build_vector into a pair of horizontal binop followed by
6455 bool isUndefLO = NumUndefsLO == Half;
6456 bool isUndefHI = NumUndefsHI == Half;
6457 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6458 isUndefLO, isUndefHI);
6462 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6463 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6465 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6466 X86Opcode = X86ISD::HADD;
6467 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6468 X86Opcode = X86ISD::HSUB;
6469 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6470 X86Opcode = X86ISD::FHADD;
6471 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6472 X86Opcode = X86ISD::FHSUB;
6476 // Don't try to expand this build_vector into a pair of horizontal add/sub
6477 // if we can simply emit a pair of scalar add/sub.
6478 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6481 // Convert this build_vector into two horizontal add/sub followed by
6483 bool isUndefLO = NumUndefsLO == Half;
6484 bool isUndefHI = NumUndefsHI == Half;
6485 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6486 isUndefLO, isUndefHI);
6493 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6496 MVT VT = Op.getSimpleValueType();
6497 MVT ExtVT = VT.getVectorElementType();
6498 unsigned NumElems = Op.getNumOperands();
6500 // Generate vectors for predicate vectors.
6501 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6502 return LowerBUILD_VECTORvXi1(Op, DAG);
6504 // Vectors containing all zeros can be matched by pxor and xorps later
6505 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6506 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6507 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6508 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6511 return getZeroVector(VT, Subtarget, DAG, dl);
6514 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6515 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6516 // vpcmpeqd on 256-bit vectors.
6517 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6518 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6521 if (!VT.is512BitVector())
6522 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
6525 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
6526 if (Broadcast.getNode())
6529 unsigned EVTBits = ExtVT.getSizeInBits();
6531 unsigned NumZero = 0;
6532 unsigned NumNonZero = 0;
6533 unsigned NonZeros = 0;
6534 bool IsAllConstants = true;
6535 SmallSet<SDValue, 8> Values;
6536 for (unsigned i = 0; i < NumElems; ++i) {
6537 SDValue Elt = Op.getOperand(i);
6538 if (Elt.getOpcode() == ISD::UNDEF)
6541 if (Elt.getOpcode() != ISD::Constant &&
6542 Elt.getOpcode() != ISD::ConstantFP)
6543 IsAllConstants = false;
6544 if (X86::isZeroNode(Elt))
6547 NonZeros |= (1 << i);
6552 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6553 if (NumNonZero == 0)
6554 return DAG.getUNDEF(VT);
6556 // Special case for single non-zero, non-undef, element.
6557 if (NumNonZero == 1) {
6558 unsigned Idx = countTrailingZeros(NonZeros);
6559 SDValue Item = Op.getOperand(Idx);
6561 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6562 // the value are obviously zero, truncate the value to i32 and do the
6563 // insertion that way. Only do this if the value is non-constant or if the
6564 // value is a constant being inserted into element 0. It is cheaper to do
6565 // a constant pool load than it is to do a movd + shuffle.
6566 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6567 (!IsAllConstants || Idx == 0)) {
6568 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6570 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6571 EVT VecVT = MVT::v4i32;
6572 unsigned VecElts = 4;
6574 // Truncate the value (which may itself be a constant) to i32, and
6575 // convert it to a vector with movd (S2V+shuffle to zero extend).
6576 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6577 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6578 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6580 // Now we have our 32-bit value zero extended in the low element of
6581 // a vector. If Idx != 0, swizzle it into place.
6583 SmallVector<int, 4> Mask;
6584 Mask.push_back(Idx);
6585 for (unsigned i = 1; i != VecElts; ++i)
6587 Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
6590 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
6594 // If we have a constant or non-constant insertion into the low element of
6595 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6596 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6597 // depending on what the source datatype is.
6600 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6602 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6603 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6604 if (VT.is256BitVector() || VT.is512BitVector()) {
6605 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6606 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6607 Item, DAG.getIntPtrConstant(0));
6609 assert(VT.is128BitVector() && "Expected an SSE value type!");
6610 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6611 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6612 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6615 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6616 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6617 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6618 if (VT.is256BitVector()) {
6619 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6620 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6622 assert(VT.is128BitVector() && "Expected an SSE value type!");
6623 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6625 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
6629 // Is it a vector logical left shift?
6630 if (NumElems == 2 && Idx == 1 &&
6631 X86::isZeroNode(Op.getOperand(0)) &&
6632 !X86::isZeroNode(Op.getOperand(1))) {
6633 unsigned NumBits = VT.getSizeInBits();
6634 return getVShift(true, VT,
6635 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6636 VT, Op.getOperand(1)),
6637 NumBits/2, DAG, *this, dl);
6640 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6643 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6644 // is a non-constant being inserted into an element other than the low one,
6645 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6646 // movd/movss) to move this into the low element, then shuffle it into
6648 if (EVTBits == 32) {
6649 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6651 // Turn it into a shuffle of zero and zero-extended scalar to vector.
6652 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
6653 SmallVector<int, 8> MaskVec;
6654 for (unsigned i = 0; i != NumElems; ++i)
6655 MaskVec.push_back(i == Idx ? 0 : 1);
6656 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
6660 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6661 if (Values.size() == 1) {
6662 if (EVTBits == 32) {
6663 // Instead of a shuffle like this:
6664 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6665 // Check if it's possible to issue this instead.
6666 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6667 unsigned Idx = countTrailingZeros(NonZeros);
6668 SDValue Item = Op.getOperand(Idx);
6669 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6670 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6675 // A vector full of immediates; various special cases are already
6676 // handled, so this is best done with a single constant-pool load.
6680 // For AVX-length vectors, build the individual 128-bit pieces and use
6681 // shuffles to put them in place.
6682 if (VT.is256BitVector() || VT.is512BitVector()) {
6683 SmallVector<SDValue, 64> V;
6684 for (unsigned i = 0; i != NumElems; ++i)
6685 V.push_back(Op.getOperand(i));
6687 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6689 // Build both the lower and upper subvector.
6690 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6691 makeArrayRef(&V[0], NumElems/2));
6692 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6693 makeArrayRef(&V[NumElems / 2], NumElems/2));
6695 // Recreate the wider vector with the lower and upper part.
6696 if (VT.is256BitVector())
6697 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6698 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6701 // Let legalizer expand 2-wide build_vectors.
6702 if (EVTBits == 64) {
6703 if (NumNonZero == 1) {
6704 // One half is zero or undef.
6705 unsigned Idx = countTrailingZeros(NonZeros);
6706 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6707 Op.getOperand(Idx));
6708 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6713 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6714 if (EVTBits == 8 && NumElems == 16) {
6715 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6717 if (V.getNode()) return V;
6720 if (EVTBits == 16 && NumElems == 8) {
6721 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6723 if (V.getNode()) return V;
6726 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6727 if (EVTBits == 32 && NumElems == 4) {
6728 SDValue V = LowerBuildVectorv4x32(Op, NumElems, NonZeros, NumNonZero,
6729 NumZero, DAG, Subtarget, *this);
6734 // If element VT is == 32 bits, turn it into a number of shuffles.
6735 SmallVector<SDValue, 8> V(NumElems);
6736 if (NumElems == 4 && NumZero > 0) {
6737 for (unsigned i = 0; i < 4; ++i) {
6738 bool isZero = !(NonZeros & (1 << i));
6740 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6742 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6745 for (unsigned i = 0; i < 2; ++i) {
6746 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6749 V[i] = V[i*2]; // Must be a zero vector.
6752 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6755 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6758 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6763 bool Reverse1 = (NonZeros & 0x3) == 2;
6764 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6768 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6769 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6771 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6774 if (Values.size() > 1 && VT.is128BitVector()) {
6775 // Check for a build vector of consecutive loads.
6776 for (unsigned i = 0; i < NumElems; ++i)
6777 V[i] = Op.getOperand(i);
6779 // Check for elements which are consecutive loads.
6780 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
6784 // Check for a build vector from mostly shuffle plus few inserting.
6785 SDValue Sh = buildFromShuffleMostly(Op, DAG);
6789 // For SSE 4.1, use insertps to put the high elements into the low element.
6790 if (getSubtarget()->hasSSE41()) {
6792 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6793 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6795 Result = DAG.getUNDEF(VT);
6797 for (unsigned i = 1; i < NumElems; ++i) {
6798 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6799 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6800 Op.getOperand(i), DAG.getIntPtrConstant(i));
6805 // Otherwise, expand into a number of unpckl*, start by extending each of
6806 // our (non-undef) elements to the full vector width with the element in the
6807 // bottom slot of the vector (which generates no code for SSE).
6808 for (unsigned i = 0; i < NumElems; ++i) {
6809 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6810 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6812 V[i] = DAG.getUNDEF(VT);
6815 // Next, we iteratively mix elements, e.g. for v4f32:
6816 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6817 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6818 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6819 unsigned EltStride = NumElems >> 1;
6820 while (EltStride != 0) {
6821 for (unsigned i = 0; i < EltStride; ++i) {
6822 // If V[i+EltStride] is undef and this is the first round of mixing,
6823 // then it is safe to just drop this shuffle: V[i] is already in the
6824 // right place, the one element (since it's the first round) being
6825 // inserted as undef can be dropped. This isn't safe for successive
6826 // rounds because they will permute elements within both vectors.
6827 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6828 EltStride == NumElems/2)
6831 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6840 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
6841 // to create 256-bit vectors from two other 128-bit ones.
6842 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6844 MVT ResVT = Op.getSimpleValueType();
6846 assert((ResVT.is256BitVector() ||
6847 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6849 SDValue V1 = Op.getOperand(0);
6850 SDValue V2 = Op.getOperand(1);
6851 unsigned NumElems = ResVT.getVectorNumElements();
6852 if(ResVT.is256BitVector())
6853 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6855 if (Op.getNumOperands() == 4) {
6856 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6857 ResVT.getVectorNumElements()/2);
6858 SDValue V3 = Op.getOperand(2);
6859 SDValue V4 = Op.getOperand(3);
6860 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6861 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6863 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6866 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6867 MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
6868 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6869 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6870 Op.getNumOperands() == 4)));
6872 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6873 // from two other 128-bit ones.
6875 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6876 return LowerAVXCONCAT_VECTORS(Op, DAG);
6880 //===----------------------------------------------------------------------===//
6881 // Vector shuffle lowering
6883 // This is an experimental code path for lowering vector shuffles on x86. It is
6884 // designed to handle arbitrary vector shuffles and blends, gracefully
6885 // degrading performance as necessary. It works hard to recognize idiomatic
6886 // shuffles and lower them to optimal instruction patterns without leaving
6887 // a framework that allows reasonably efficient handling of all vector shuffle
6889 //===----------------------------------------------------------------------===//
6891 /// \brief Tiny helper function to identify a no-op mask.
6893 /// This is a somewhat boring predicate function. It checks whether the mask
6894 /// array input, which is assumed to be a single-input shuffle mask of the kind
6895 /// used by the X86 shuffle instructions (not a fully general
6896 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6897 /// in-place shuffle are 'no-op's.
6898 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6899 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6900 if (Mask[i] != -1 && Mask[i] != i)
6905 /// \brief Helper function to classify a mask as a single-input mask.
6907 /// This isn't a generic single-input test because in the vector shuffle
6908 /// lowering we canonicalize single inputs to be the first input operand. This
6909 /// means we can more quickly test for a single input by only checking whether
6910 /// an input from the second operand exists. We also assume that the size of
6911 /// mask corresponds to the size of the input vectors which isn't true in the
6912 /// fully general case.
6913 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6915 if (M >= (int)Mask.size())
6920 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6922 /// This helper function produces an 8-bit shuffle immediate corresponding to
6923 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6924 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6927 /// NB: We rely heavily on "undef" masks preserving the input lane.
6928 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
6929 SelectionDAG &DAG) {
6930 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6931 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6932 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6933 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6934 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6937 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6938 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6939 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6940 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6941 return DAG.getConstant(Imm, MVT::i8);
6944 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
6946 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
6947 /// support for floating point shuffles but not integer shuffles. These
6948 /// instructions will incur a domain crossing penalty on some chips though so
6949 /// it is better to avoid lowering through this for integer vectors where
6951 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
6952 const X86Subtarget *Subtarget,
6953 SelectionDAG &DAG) {
6955 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
6956 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
6957 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
6958 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6959 ArrayRef<int> Mask = SVOp->getMask();
6960 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
6962 if (isSingleInputShuffleMask(Mask)) {
6963 // Straight shuffle of a single input vector. Simulate this by using the
6964 // single input as both of the "inputs" to this instruction..
6965 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
6966 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
6967 DAG.getConstant(SHUFPDMask, MVT::i8));
6969 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
6970 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
6972 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
6973 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
6974 DAG.getConstant(SHUFPDMask, MVT::i8));
6977 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
6979 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
6980 /// the integer unit to minimize domain crossing penalties. However, for blends
6981 /// it falls back to the floating point shuffle operation with appropriate bit
6983 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
6984 const X86Subtarget *Subtarget,
6985 SelectionDAG &DAG) {
6987 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
6988 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
6989 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
6990 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6991 ArrayRef<int> Mask = SVOp->getMask();
6992 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
6994 if (isSingleInputShuffleMask(Mask)) {
6995 // Straight shuffle of a single input vector. For everything from SSE2
6996 // onward this has a single fast instruction with no scary immediates.
6997 // We have to map the mask as it is actually a v4i32 shuffle instruction.
6998 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
6999 int WidenedMask[4] = {
7000 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7001 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7003 ISD::BITCAST, DL, MVT::v2i64,
7004 DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
7005 getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
7008 // We implement this with SHUFPD which is pretty lame because it will likely
7009 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7010 // However, all the alternatives are still more cycles and newer chips don't
7011 // have this problem. It would be really nice if x86 had better shuffles here.
7012 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
7013 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
7014 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7015 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7018 /// \brief Lower 4-lane 32-bit floating point shuffles.
7020 /// Uses instructions exclusively from the floating point unit to minimize
7021 /// domain crossing penalties, as these are sufficient to implement all v4f32
7023 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7024 const X86Subtarget *Subtarget,
7025 SelectionDAG &DAG) {
7027 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7028 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7029 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7030 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7031 ArrayRef<int> Mask = SVOp->getMask();
7032 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7034 SDValue LowV = V1, HighV = V2;
7035 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7038 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7040 if (NumV2Elements == 0)
7041 // Straight shuffle of a single input vector. We pass the input vector to
7042 // both operands to simulate this with a SHUFPS.
7043 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
7044 getV4X86ShuffleImm8ForMask(Mask, DAG));
7046 if (NumV2Elements == 1) {
7048 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7050 // Compute the index adjacent to V2Index and in the same half by toggling
7052 int V2AdjIndex = V2Index ^ 1;
7054 if (Mask[V2AdjIndex] == -1) {
7055 // Handles all the cases where we have a single V2 element and an undef.
7056 // This will only ever happen in the high lanes because we commute the
7057 // vector otherwise.
7059 std::swap(LowV, HighV);
7060 NewMask[V2Index] -= 4;
7062 // Handle the case where the V2 element ends up adjacent to a V1 element.
7063 // To make this work, blend them together as the first step.
7064 int V1Index = V2AdjIndex;
7065 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7066 V2 = DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V2, V1,
7067 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7069 // Now proceed to reconstruct the final blend as we have the necessary
7070 // high or low half formed.
7077 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7078 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7080 } else if (NumV2Elements == 2) {
7081 if (Mask[0] < 4 && Mask[1] < 4) {
7082 // Handle the easy case where we have V1 in the low lanes and V2 in the
7083 // high lanes. We never see this reversed because we sort the shuffle.
7087 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7088 // trying to place elements directly, just blend them and set up the final
7089 // shuffle to place them.
7091 // The first two blend mask elements are for V1, the second two are for
7093 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
7094 Mask[2] < 4 ? Mask[2] : Mask[3],
7095 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
7096 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
7097 V1 = DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V2,
7098 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7100 // Now we do a normal shuffle of V1 by giving V1 as both operands to
7103 NewMask[0] = Mask[0] < 4 ? 0 : 2;
7104 NewMask[1] = Mask[0] < 4 ? 2 : 0;
7105 NewMask[2] = Mask[2] < 4 ? 1 : 3;
7106 NewMask[3] = Mask[2] < 4 ? 3 : 1;
7109 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, LowV, HighV,
7110 getV4X86ShuffleImm8ForMask(NewMask, DAG));
7113 /// \brief Lower 4-lane i32 vector shuffles.
7115 /// We try to handle these with integer-domain shuffles where we can, but for
7116 /// blends we use the floating point domain blend instructions.
7117 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7118 const X86Subtarget *Subtarget,
7119 SelectionDAG &DAG) {
7121 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
7122 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7123 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7124 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7125 ArrayRef<int> Mask = SVOp->getMask();
7126 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7128 if (isSingleInputShuffleMask(Mask))
7129 // Straight shuffle of a single input vector. For everything from SSE2
7130 // onward this has a single fast instruction with no scary immediates.
7131 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7132 getV4X86ShuffleImm8ForMask(Mask, DAG));
7134 // We implement this with SHUFPS because it can blend from two vectors.
7135 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
7136 // up the inputs, bypassing domain shift penalties that we would encur if we
7137 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
7139 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
7140 DAG.getVectorShuffle(
7142 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
7143 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
7146 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
7147 /// shuffle lowering, and the most complex part.
7149 /// The lowering strategy is to try to form pairs of input lanes which are
7150 /// targeted at the same half of the final vector, and then use a dword shuffle
7151 /// to place them onto the right half, and finally unpack the paired lanes into
7152 /// their final position.
7154 /// The exact breakdown of how to form these dword pairs and align them on the
7155 /// correct sides is really tricky. See the comments within the function for
7156 /// more of the details.
7157 static SDValue lowerV8I16SingleInputVectorShuffle(
7158 SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
7159 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7160 assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
7161 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
7162 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
7164 SmallVector<int, 4> LoInputs;
7165 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
7166 [](int M) { return M >= 0; });
7167 std::sort(LoInputs.begin(), LoInputs.end());
7168 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
7169 SmallVector<int, 4> HiInputs;
7170 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
7171 [](int M) { return M >= 0; });
7172 std::sort(HiInputs.begin(), HiInputs.end());
7173 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
7175 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
7176 int NumHToL = LoInputs.size() - NumLToL;
7178 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
7179 int NumHToH = HiInputs.size() - NumLToH;
7180 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
7181 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
7182 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
7183 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
7185 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
7186 // such inputs we can swap two of the dwords across the half mark and end up
7187 // with <=2 inputs to each half in each half. Once there, we can fall through
7188 // to the generic code below. For example:
7190 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7191 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
7193 // Before we had 3-1 in the low half and 3-1 in the high half. Afterward, 2-2
7195 auto balanceSides = [&](ArrayRef<int> ThreeInputs, int OneInput,
7196 int ThreeInputHalfSum, int OneInputHalfOffset) {
7197 // Compute the index of dword with only one word among the three inputs in
7198 // a half by taking the sum of the half with three inputs and subtracting
7199 // the sum of the actual three inputs. The difference is the remaining
7201 int DWordA = (ThreeInputHalfSum -
7202 std::accumulate(ThreeInputs.begin(), ThreeInputs.end(), 0)) /
7204 int DWordB = OneInputHalfOffset / 2 + (OneInput / 2 + 1) % 2;
7206 int PSHUFDMask[] = {0, 1, 2, 3};
7207 PSHUFDMask[DWordA] = DWordB;
7208 PSHUFDMask[DWordB] = DWordA;
7209 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
7210 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7211 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
7212 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
7214 // Adjust the mask to match the new locations of A and B.
7216 if (M != -1 && M/2 == DWordA)
7217 M = 2 * DWordB + M % 2;
7218 else if (M != -1 && M/2 == DWordB)
7219 M = 2 * DWordA + M % 2;
7221 // Recurse back into this routine to re-compute state now that this isn't
7222 // a 3 and 1 problem.
7223 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
7226 if (NumLToL == 3 && NumHToL == 1)
7227 return balanceSides(LToLInputs, HToLInputs[0], 0 + 1 + 2 + 3, 4);
7228 else if (NumLToL == 1 && NumHToL == 3)
7229 return balanceSides(HToLInputs, LToLInputs[0], 4 + 5 + 6 + 7, 0);
7230 else if (NumLToH == 1 && NumHToH == 3)
7231 return balanceSides(HToHInputs, LToHInputs[0], 4 + 5 + 6 + 7, 0);
7232 else if (NumLToH == 3 && NumHToH == 1)
7233 return balanceSides(LToHInputs, HToHInputs[0], 0 + 1 + 2 + 3, 4);
7235 // At this point there are at most two inputs to the low and high halves from
7236 // each half. That means the inputs can always be grouped into dwords and
7237 // those dwords can then be moved to the correct half with a dword shuffle.
7238 // We use at most one low and one high word shuffle to collect these paired
7239 // inputs into dwords, and finally a dword shuffle to place them.
7240 int PSHUFLMask[4] = {-1, -1, -1, -1};
7241 int PSHUFHMask[4] = {-1, -1, -1, -1};
7242 int PSHUFDMask[4] = {-1, -1, -1, -1};
7244 // First fix the masks for all the inputs that are staying in their
7245 // original halves. This will then dictate the targets of the cross-half
7247 auto fixInPlaceInputs = [&PSHUFDMask](
7248 ArrayRef<int> InPlaceInputs, MutableArrayRef<int> SourceHalfMask,
7249 MutableArrayRef<int> HalfMask, int HalfOffset) {
7250 if (InPlaceInputs.empty())
7252 if (InPlaceInputs.size() == 1) {
7253 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
7254 InPlaceInputs[0] - HalfOffset;
7255 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
7259 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
7260 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
7261 InPlaceInputs[0] - HalfOffset;
7262 // Put the second input next to the first so that they are packed into
7263 // a dword. We find the adjacent index by toggling the low bit.
7264 int AdjIndex = InPlaceInputs[0] ^ 1;
7265 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
7266 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
7267 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
7269 if (!HToLInputs.empty())
7270 fixInPlaceInputs(LToLInputs, PSHUFLMask, LoMask, 0);
7271 if (!LToHInputs.empty())
7272 fixInPlaceInputs(HToHInputs, PSHUFHMask, HiMask, 4);
7274 // Now gather the cross-half inputs and place them into a free dword of
7275 // their target half.
7276 // FIXME: This operation could almost certainly be simplified dramatically to
7277 // look more like the 3-1 fixing operation.
7278 auto moveInputsToRightHalf = [&PSHUFDMask](
7279 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
7280 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
7281 int SourceOffset, int DestOffset) {
7282 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
7283 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
7285 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
7287 int LowWord = Word & ~1;
7288 int HighWord = Word | 1;
7289 return isWordClobbered(SourceHalfMask, LowWord) ||
7290 isWordClobbered(SourceHalfMask, HighWord);
7293 if (IncomingInputs.empty())
7296 if (ExistingInputs.empty()) {
7297 // Map any dwords with inputs from them into the right half.
7298 for (int Input : IncomingInputs) {
7299 // If the source half mask maps over the inputs, turn those into
7300 // swaps and use the swapped lane.
7301 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
7302 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
7303 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
7304 Input - SourceOffset;
7305 // We have to swap the uses in our half mask in one sweep.
7306 for (int &M : HalfMask)
7307 if (M == SourceHalfMask[Input - SourceOffset])
7309 else if (M == Input)
7310 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
7312 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
7313 Input - SourceOffset &&
7314 "Previous placement doesn't match!");
7316 // Note that this correctly re-maps both when we do a swap and when
7317 // we observe the other side of the swap above. We rely on that to
7318 // avoid swapping the members of the input list directly.
7319 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
7322 // Map the input's dword into the correct half.
7323 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
7324 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
7326 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
7328 "Previous placement doesn't match!");
7331 // And just directly shift any other-half mask elements to be same-half
7332 // as we will have mirrored the dword containing the element into the
7333 // same position within that half.
7334 for (int &M : HalfMask)
7335 if (M >= SourceOffset && M < SourceOffset + 4) {
7336 M = M - SourceOffset + DestOffset;
7337 assert(M >= 0 && "This should never wrap below zero!");
7342 // Ensure we have the input in a viable dword of its current half. This
7343 // is particularly tricky because the original position may be clobbered
7344 // by inputs being moved and *staying* in that half.
7345 if (IncomingInputs.size() == 1) {
7346 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
7347 int InputFixed = std::find(std::begin(SourceHalfMask),
7348 std::end(SourceHalfMask), -1) -
7349 std::begin(SourceHalfMask) + SourceOffset;
7350 SourceHalfMask[InputFixed - SourceOffset] =
7351 IncomingInputs[0] - SourceOffset;
7352 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
7354 IncomingInputs[0] = InputFixed;
7356 } else if (IncomingInputs.size() == 2) {
7357 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
7358 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
7359 int SourceDWordBase = !isDWordClobbered(SourceHalfMask, 0) ? 0 : 2;
7360 assert(!isDWordClobbered(SourceHalfMask, SourceDWordBase) &&
7361 "Not all dwords can be clobbered!");
7362 SourceHalfMask[SourceDWordBase] = IncomingInputs[0] - SourceOffset;
7363 SourceHalfMask[SourceDWordBase + 1] = IncomingInputs[1] - SourceOffset;
7364 for (int &M : HalfMask)
7365 if (M == IncomingInputs[0])
7366 M = SourceDWordBase + SourceOffset;
7367 else if (M == IncomingInputs[1])
7368 M = SourceDWordBase + 1 + SourceOffset;
7369 IncomingInputs[0] = SourceDWordBase + SourceOffset;
7370 IncomingInputs[1] = SourceDWordBase + 1 + SourceOffset;
7373 llvm_unreachable("Unhandled input size!");
7376 // Now hoist the DWord down to the right half.
7377 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
7378 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
7379 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
7380 for (int Input : IncomingInputs)
7381 std::replace(HalfMask.begin(), HalfMask.end(), Input,
7382 FreeDWord * 2 + Input % 2);
7384 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask,
7385 /*SourceOffset*/ 4, /*DestOffset*/ 0);
7386 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask,
7387 /*SourceOffset*/ 0, /*DestOffset*/ 4);
7389 // Now enact all the shuffles we've computed to move the inputs into their
7391 if (!isNoopShuffleMask(PSHUFLMask))
7392 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
7393 getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
7394 if (!isNoopShuffleMask(PSHUFHMask))
7395 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
7396 getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
7397 if (!isNoopShuffleMask(PSHUFDMask))
7398 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
7399 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7400 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
7401 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
7403 // At this point, each half should contain all its inputs, and we can then
7404 // just shuffle them into their final position.
7405 assert(std::count_if(LoMask.begin(), LoMask.end(),
7406 [](int M) { return M >= 4; }) == 0 &&
7407 "Failed to lift all the high half inputs to the low mask!");
7408 assert(std::count_if(HiMask.begin(), HiMask.end(),
7409 [](int M) { return M >= 0 && M < 4; }) == 0 &&
7410 "Failed to lift all the low half inputs to the high mask!");
7412 // Do a half shuffle for the low mask.
7413 if (!isNoopShuffleMask(LoMask))
7414 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
7415 getV4X86ShuffleImm8ForMask(LoMask, DAG));
7417 // Do a half shuffle with the high mask after shifting its values down.
7418 for (int &M : HiMask)
7421 if (!isNoopShuffleMask(HiMask))
7422 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
7423 getV4X86ShuffleImm8ForMask(HiMask, DAG));
7428 /// \brief Detect whether the mask pattern should be lowered through
7431 /// This essentially tests whether viewing the mask as an interleaving of two
7432 /// sub-sequences reduces the cross-input traffic of a blend operation. If so,
7433 /// lowering it through interleaving is a significantly better strategy.
7434 static bool shouldLowerAsInterleaving(ArrayRef<int> Mask) {
7435 int NumEvenInputs[2] = {0, 0};
7436 int NumOddInputs[2] = {0, 0};
7437 int NumLoInputs[2] = {0, 0};
7438 int NumHiInputs[2] = {0, 0};
7439 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7443 int InputIdx = Mask[i] >= Size;
7446 ++NumLoInputs[InputIdx];
7448 ++NumHiInputs[InputIdx];
7451 ++NumEvenInputs[InputIdx];
7453 ++NumOddInputs[InputIdx];
7456 // The minimum number of cross-input results for both the interleaved and
7457 // split cases. If interleaving results in fewer cross-input results, return
7459 int InterleavedCrosses = std::min(NumEvenInputs[1] + NumOddInputs[0],
7460 NumEvenInputs[0] + NumOddInputs[1]);
7461 int SplitCrosses = std::min(NumLoInputs[1] + NumHiInputs[0],
7462 NumLoInputs[0] + NumHiInputs[1]);
7463 return InterleavedCrosses < SplitCrosses;
7466 /// \brief Blend two v8i16 vectors using a naive unpack strategy.
7468 /// This strategy only works when the inputs from each vector fit into a single
7469 /// half of that vector, and generally there are not so many inputs as to leave
7470 /// the in-place shuffles required highly constrained (and thus expensive). It
7471 /// shifts all the inputs into a single side of both input vectors and then
7472 /// uses an unpack to interleave these inputs in a single vector. At that
7473 /// point, we will fall back on the generic single input shuffle lowering.
7474 static SDValue lowerV8I16BasicBlendVectorShuffle(SDLoc DL, SDValue V1,
7476 MutableArrayRef<int> Mask,
7477 const X86Subtarget *Subtarget,
7478 SelectionDAG &DAG) {
7479 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
7480 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
7481 SmallVector<int, 3> LoV1Inputs, HiV1Inputs, LoV2Inputs, HiV2Inputs;
7482 for (int i = 0; i < 8; ++i)
7483 if (Mask[i] >= 0 && Mask[i] < 4)
7484 LoV1Inputs.push_back(i);
7485 else if (Mask[i] >= 4 && Mask[i] < 8)
7486 HiV1Inputs.push_back(i);
7487 else if (Mask[i] >= 8 && Mask[i] < 12)
7488 LoV2Inputs.push_back(i);
7489 else if (Mask[i] >= 12)
7490 HiV2Inputs.push_back(i);
7492 int NumV1Inputs = LoV1Inputs.size() + HiV1Inputs.size();
7493 int NumV2Inputs = LoV2Inputs.size() + HiV2Inputs.size();
7496 assert(NumV1Inputs > 0 && NumV1Inputs <= 3 && "At most 3 inputs supported");
7497 assert(NumV2Inputs > 0 && NumV2Inputs <= 3 && "At most 3 inputs supported");
7498 assert(NumV1Inputs + NumV2Inputs <= 4 && "At most 4 combined inputs");
7500 bool MergeFromLo = LoV1Inputs.size() + LoV2Inputs.size() >=
7501 HiV1Inputs.size() + HiV2Inputs.size();
7503 auto moveInputsToHalf = [&](SDValue V, ArrayRef<int> LoInputs,
7504 ArrayRef<int> HiInputs, bool MoveToLo,
7506 ArrayRef<int> GoodInputs = MoveToLo ? LoInputs : HiInputs;
7507 ArrayRef<int> BadInputs = MoveToLo ? HiInputs : LoInputs;
7508 if (BadInputs.empty())
7511 int MoveMask[] = {-1, -1, -1, -1, -1, -1, -1, -1};
7512 int MoveOffset = MoveToLo ? 0 : 4;
7514 if (GoodInputs.empty()) {
7515 for (int BadInput : BadInputs) {
7516 MoveMask[Mask[BadInput] % 4 + MoveOffset] = Mask[BadInput] - MaskOffset;
7517 Mask[BadInput] = Mask[BadInput] % 4 + MoveOffset + MaskOffset;
7520 if (GoodInputs.size() == 2) {
7521 // If the low inputs are spread across two dwords, pack them into
7523 MoveMask[Mask[GoodInputs[0]] % 2 + MoveOffset] =
7524 Mask[GoodInputs[0]] - MaskOffset;
7525 MoveMask[Mask[GoodInputs[1]] % 2 + MoveOffset] =
7526 Mask[GoodInputs[1]] - MaskOffset;
7527 Mask[GoodInputs[0]] = Mask[GoodInputs[0]] % 2 + MoveOffset + MaskOffset;
7528 Mask[GoodInputs[1]] = Mask[GoodInputs[0]] % 2 + MoveOffset + MaskOffset;
7530 // Otherwise pin the low inputs.
7531 for (int GoodInput : GoodInputs)
7532 MoveMask[Mask[GoodInput] - MaskOffset] = Mask[GoodInput] - MaskOffset;
7536 std::find(std::begin(MoveMask) + MoveOffset, std::end(MoveMask), -1) -
7537 std::begin(MoveMask);
7538 assert(MoveMaskIdx >= MoveOffset && "Established above");
7540 if (BadInputs.size() == 2) {
7541 assert(MoveMask[MoveMaskIdx] == -1 && "Expected empty slot");
7542 assert(MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot");
7543 MoveMask[MoveMaskIdx + Mask[BadInputs[0]] % 2] =
7544 Mask[BadInputs[0]] - MaskOffset;
7545 MoveMask[MoveMaskIdx + Mask[BadInputs[1]] % 2] =
7546 Mask[BadInputs[1]] - MaskOffset;
7547 Mask[BadInputs[0]] = MoveMaskIdx + Mask[BadInputs[0]] % 2 + MaskOffset;
7548 Mask[BadInputs[1]] = MoveMaskIdx + Mask[BadInputs[1]] % 2 + MaskOffset;
7550 assert(BadInputs.size() == 1 && "All sizes handled");
7551 MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
7552 Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
7556 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
7559 V1 = moveInputsToHalf(V1, LoV1Inputs, HiV1Inputs, MergeFromLo,
7561 V2 = moveInputsToHalf(V2, LoV2Inputs, HiV2Inputs, MergeFromLo,
7564 // FIXME: Select an interleaving of the merge of V1 and V2 that minimizes
7565 // cross-half traffic in the final shuffle.
7567 // Munge the mask to be a single-input mask after the unpack merges the
7571 M = 2 * (M % 4) + (M / 8);
7573 return DAG.getVectorShuffle(
7574 MVT::v8i16, DL, DAG.getNode(MergeFromLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
7575 DL, MVT::v8i16, V1, V2),
7576 DAG.getUNDEF(MVT::v8i16), Mask);
7579 /// \brief Generic lowering of 8-lane i16 shuffles.
7581 /// This handles both single-input shuffles and combined shuffle/blends with
7582 /// two inputs. The single input shuffles are immediately delegated to
7583 /// a dedicated lowering routine.
7585 /// The blends are lowered in one of three fundamental ways. If there are few
7586 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
7587 /// of the input is significantly cheaper when lowered as an interleaving of
7588 /// the two inputs, try to interleave them. Otherwise, blend the low and high
7589 /// halves of the inputs separately (making them have relatively few inputs)
7590 /// and then concatenate them.
7591 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7592 const X86Subtarget *Subtarget,
7593 SelectionDAG &DAG) {
7595 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
7596 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
7597 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
7598 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7599 ArrayRef<int> OrigMask = SVOp->getMask();
7600 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
7601 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
7602 MutableArrayRef<int> Mask(MaskStorage);
7604 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
7606 auto isV1 = [](int M) { return M >= 0 && M < 8; };
7607 auto isV2 = [](int M) { return M >= 8; };
7609 int NumV1Inputs = std::count_if(Mask.begin(), Mask.end(), isV1);
7610 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
7612 if (NumV2Inputs == 0)
7613 return lowerV8I16SingleInputVectorShuffle(DL, V1, Mask, Subtarget, DAG);
7615 assert(NumV1Inputs > 0 && "All single-input shuffles should be canonicalized "
7616 "to be V1-input shuffles.");
7618 if (NumV1Inputs + NumV2Inputs <= 4)
7619 return lowerV8I16BasicBlendVectorShuffle(DL, V1, V2, Mask, Subtarget, DAG);
7621 // Check whether an interleaving lowering is likely to be more efficient.
7622 // This isn't perfect but it is a strong heuristic that tends to work well on
7623 // the kinds of shuffles that show up in practice.
7625 // FIXME: Handle 1x, 2x, and 4x interleaving.
7626 if (shouldLowerAsInterleaving(Mask)) {
7627 // FIXME: Figure out whether we should pack these into the low or high
7630 int EMask[8], OMask[8];
7631 for (int i = 0; i < 4; ++i) {
7632 EMask[i] = Mask[2*i];
7633 OMask[i] = Mask[2*i + 1];
7638 SDValue Evens = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, EMask);
7639 SDValue Odds = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, OMask);
7641 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, Evens, Odds);
7644 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
7645 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
7647 for (int i = 0; i < 4; ++i) {
7648 LoBlendMask[i] = Mask[i];
7649 HiBlendMask[i] = Mask[i + 4];
7652 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
7653 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
7654 LoV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, LoV);
7655 HiV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, HiV);
7657 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
7658 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, LoV, HiV));
7661 /// \brief Generic lowering of v16i8 shuffles.
7663 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
7664 /// detect any complexity reducing interleaving. If that doesn't help, it uses
7665 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
7666 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
7668 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7669 const X86Subtarget *Subtarget,
7670 SelectionDAG &DAG) {
7672 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
7673 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
7674 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
7675 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7676 ArrayRef<int> OrigMask = SVOp->getMask();
7677 assert(OrigMask.size() == 16 && "Unexpected mask size for v16 shuffle!");
7678 int MaskStorage[16] = {
7679 OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
7680 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7],
7681 OrigMask[8], OrigMask[9], OrigMask[10], OrigMask[11],
7682 OrigMask[12], OrigMask[13], OrigMask[14], OrigMask[15]};
7683 MutableArrayRef<int> Mask(MaskStorage);
7684 MutableArrayRef<int> LoMask = Mask.slice(0, 8);
7685 MutableArrayRef<int> HiMask = Mask.slice(8, 8);
7687 // For single-input shuffles, there are some nicer lowering tricks we can use.
7688 if (isSingleInputShuffleMask(Mask)) {
7689 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
7690 // Notably, this handles splat and partial-splat shuffles more efficiently.
7691 // However, it only makes sense if the pre-duplication shuffle simplifies
7692 // things significantly. Currently, this means we need to be able to
7693 // express the pre-duplication shuffle as an i16 shuffle.
7695 // FIXME: We should check for other patterns which can be widened into an
7696 // i16 shuffle as well.
7697 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
7698 for (int i = 0; i < 16; i += 2) {
7699 if (Mask[i] != Mask[i + 1])
7704 auto tryToWidenViaDuplication = [&]() -> SDValue {
7705 if (!canWidenViaDuplication(Mask))
7707 SmallVector<int, 4> LoInputs;
7708 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
7709 [](int M) { return M >= 0 && M < 8; });
7710 std::sort(LoInputs.begin(), LoInputs.end());
7711 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
7713 SmallVector<int, 4> HiInputs;
7714 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
7715 [](int M) { return M >= 8; });
7716 std::sort(HiInputs.begin(), HiInputs.end());
7717 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
7720 bool TargetLo = LoInputs.size() >= HiInputs.size();
7721 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
7722 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
7724 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
7725 SmallDenseMap<int, int, 8> LaneMap;
7726 for (int I : InPlaceInputs) {
7727 PreDupI16Shuffle[I/2] = I/2;
7730 int j = TargetLo ? 0 : 4, je = j + 4;
7731 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
7732 // Check if j is already a shuffle of this input. This happens when
7733 // there are two adjacent bytes after we move the low one.
7734 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
7735 // If we haven't yet mapped the input, search for a slot into which
7737 while (j < je && PreDupI16Shuffle[j] != -1)
7741 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
7744 // Map this input with the i16 shuffle.
7745 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
7748 // Update the lane map based on the mapping we ended up with.
7749 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
7752 ISD::BITCAST, DL, MVT::v16i8,
7753 DAG.getVectorShuffle(MVT::v8i16, DL,
7754 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
7755 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
7757 // Unpack the bytes to form the i16s that will be shuffled into place.
7758 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
7759 MVT::v16i8, V1, V1);
7761 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
7762 for (int i = 0; i < 16; i += 2) {
7764 PostDupI16Shuffle[i / 2] = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
7765 assert(PostDupI16Shuffle[i / 2] < 8 && "Invalid v8 shuffle mask!");
7768 ISD::BITCAST, DL, MVT::v16i8,
7769 DAG.getVectorShuffle(MVT::v8i16, DL,
7770 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
7771 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
7773 if (SDValue V = tryToWidenViaDuplication())
7777 // Check whether an interleaving lowering is likely to be more efficient.
7778 // This isn't perfect but it is a strong heuristic that tends to work well on
7779 // the kinds of shuffles that show up in practice.
7781 // FIXME: We need to handle other interleaving widths (i16, i32, ...).
7782 if (shouldLowerAsInterleaving(Mask)) {
7783 // FIXME: Figure out whether we should pack these into the low or high
7786 int EMask[16], OMask[16];
7787 for (int i = 0; i < 8; ++i) {
7788 EMask[i] = Mask[2*i];
7789 OMask[i] = Mask[2*i + 1];
7794 SDValue Evens = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, EMask);
7795 SDValue Odds = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, OMask);
7797 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, Evens, Odds);
7800 int V1LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
7801 int V1HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
7802 int V2LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
7803 int V2HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
7805 auto buildBlendMasks = [](MutableArrayRef<int> HalfMask,
7806 MutableArrayRef<int> V1HalfBlendMask,
7807 MutableArrayRef<int> V2HalfBlendMask) {
7808 for (int i = 0; i < 8; ++i)
7809 if (HalfMask[i] >= 0 && HalfMask[i] < 16) {
7810 V1HalfBlendMask[i] = HalfMask[i];
7812 } else if (HalfMask[i] >= 16) {
7813 V2HalfBlendMask[i] = HalfMask[i] - 16;
7814 HalfMask[i] = i + 8;
7817 buildBlendMasks(LoMask, V1LoBlendMask, V2LoBlendMask);
7818 buildBlendMasks(HiMask, V1HiBlendMask, V2HiBlendMask);
7820 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
7822 auto buildLoAndHiV8s = [&](SDValue V, MutableArrayRef<int> LoBlendMask,
7823 MutableArrayRef<int> HiBlendMask) {
7825 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
7826 // them out and avoid using UNPCK{L,H} to extract the elements of V as
7828 if (std::none_of(LoBlendMask.begin(), LoBlendMask.end(),
7829 [](int M) { return M >= 0 && M % 2 == 1; }) &&
7830 std::none_of(HiBlendMask.begin(), HiBlendMask.end(),
7831 [](int M) { return M >= 0 && M % 2 == 1; })) {
7832 // Use a mask to drop the high bytes.
7833 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
7834 V1 = DAG.getNode(ISD::AND, DL, MVT::v8i16, V1,
7835 DAG.getConstant(0x00FF, MVT::v8i16));
7837 // This will be a single vector shuffle instead of a blend so nuke V2.
7838 V2 = DAG.getUNDEF(MVT::v8i16);
7840 // Squash the masks to point directly into V1.
7841 for (int &M : LoBlendMask)
7844 for (int &M : HiBlendMask)
7848 // Otherwise just unpack the low half of V into V1 and the high half into
7849 // V2 so that we can blend them as i16s.
7850 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
7851 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
7852 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
7853 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
7856 SDValue BlendedLo = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
7857 SDValue BlendedHi = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
7858 return std::make_pair(BlendedLo, BlendedHi);
7860 SDValue V1Lo, V1Hi, V2Lo, V2Hi;
7861 std::tie(V1Lo, V1Hi) = buildLoAndHiV8s(V1, V1LoBlendMask, V1HiBlendMask);
7862 std::tie(V2Lo, V2Hi) = buildLoAndHiV8s(V2, V2LoBlendMask, V2HiBlendMask);
7864 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Lo, V2Lo, LoMask);
7865 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Hi, V2Hi, HiMask);
7867 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
7870 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
7872 /// This routine breaks down the specific type of 128-bit shuffle and
7873 /// dispatches to the lowering routines accordingly.
7874 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7875 MVT VT, const X86Subtarget *Subtarget,
7876 SelectionDAG &DAG) {
7877 switch (VT.SimpleTy) {
7879 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
7881 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
7883 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
7885 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
7887 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
7889 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
7892 llvm_unreachable("Unimplemented!");
7896 /// \brief Tiny helper function to test whether adjacent masks are sequential.
7897 static bool areAdjacentMasksSequential(ArrayRef<int> Mask) {
7898 for (int i = 0, Size = Mask.size(); i < Size; i += 2)
7899 if (Mask[i] + 1 != Mask[i+1])
7905 /// \brief Top-level lowering for x86 vector shuffles.
7907 /// This handles decomposition, canonicalization, and lowering of all x86
7908 /// vector shuffles. Most of the specific lowering strategies are encapsulated
7909 /// above in helper routines. The canonicalization attempts to widen shuffles
7910 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
7911 /// s.t. only one of the two inputs needs to be tested, etc.
7912 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
7913 SelectionDAG &DAG) {
7914 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7915 ArrayRef<int> Mask = SVOp->getMask();
7916 SDValue V1 = Op.getOperand(0);
7917 SDValue V2 = Op.getOperand(1);
7918 MVT VT = Op.getSimpleValueType();
7919 int NumElements = VT.getVectorNumElements();
7922 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
7924 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
7925 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
7926 if (V1IsUndef && V2IsUndef)
7927 return DAG.getUNDEF(VT);
7929 // When we create a shuffle node we put the UNDEF node to second operand,
7930 // but in some cases the first operand may be transformed to UNDEF.
7931 // In this case we should just commute the node.
7933 return DAG.getCommutedVectorShuffle(*SVOp);
7935 // Check for non-undef masks pointing at an undef vector and make the masks
7936 // undef as well. This makes it easier to match the shuffle based solely on
7940 if (M >= NumElements) {
7941 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
7942 for (int &M : NewMask)
7943 if (M >= NumElements)
7945 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
7948 // For integer vector shuffles, try to collapse them into a shuffle of fewer
7949 // lanes but wider integers. We cap this to not form integers larger than i64
7950 // but it might be interesting to form i128 integers to handle flipping the
7951 // low and high halves of AVX 256-bit vectors.
7952 if (VT.isInteger() && VT.getScalarSizeInBits() < 64 &&
7953 areAdjacentMasksSequential(Mask)) {
7954 SmallVector<int, 8> NewMask;
7955 for (int i = 0, Size = Mask.size(); i < Size; i += 2)
7956 NewMask.push_back(Mask[i] / 2);
7958 MVT::getVectorVT(MVT::getIntegerVT(VT.getScalarSizeInBits() * 2),
7959 VT.getVectorNumElements() / 2);
7960 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
7961 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
7962 return DAG.getNode(ISD::BITCAST, dl, VT,
7963 DAG.getVectorShuffle(NewVT, dl, V1, V2, NewMask));
7966 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
7967 for (int M : SVOp->getMask())
7970 else if (M < NumElements)
7975 // Commute the shuffle as needed such that more elements come from V1 than
7976 // V2. This allows us to match the shuffle pattern strictly on how many
7977 // elements come from V1 without handling the symmetric cases.
7978 if (NumV2Elements > NumV1Elements)
7979 return DAG.getCommutedVectorShuffle(*SVOp);
7981 // When the number of V1 and V2 elements are the same, try to minimize the
7982 // number of uses of V2 in the low half of the vector.
7983 if (NumV1Elements == NumV2Elements) {
7984 int LowV1Elements = 0, LowV2Elements = 0;
7985 for (int M : SVOp->getMask().slice(0, NumElements / 2))
7986 if (M >= NumElements)
7990 if (LowV2Elements > LowV1Elements)
7991 return DAG.getCommutedVectorShuffle(*SVOp);
7994 // For each vector width, delegate to a specialized lowering routine.
7995 if (VT.getSizeInBits() == 128)
7996 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
7998 llvm_unreachable("Unimplemented!");
8002 //===----------------------------------------------------------------------===//
8003 // Legacy vector shuffle lowering
8005 // This code is the legacy code handling vector shuffles until the above
8006 // replaces its functionality and performance.
8007 //===----------------------------------------------------------------------===//
8009 static bool isBlendMask(ArrayRef<int> MaskVals, MVT VT, bool hasSSE41,
8010 bool hasInt256, unsigned *MaskOut = nullptr) {
8011 MVT EltVT = VT.getVectorElementType();
8013 // There is no blend with immediate in AVX-512.
8014 if (VT.is512BitVector())
8017 if (!hasSSE41 || EltVT == MVT::i8)
8019 if (!hasInt256 && VT == MVT::v16i16)
8022 unsigned MaskValue = 0;
8023 unsigned NumElems = VT.getVectorNumElements();
8024 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
8025 unsigned NumLanes = (NumElems - 1) / 8 + 1;
8026 unsigned NumElemsInLane = NumElems / NumLanes;
8028 // Blend for v16i16 should be symetric for the both lanes.
8029 for (unsigned i = 0; i < NumElemsInLane; ++i) {
8031 int SndLaneEltIdx = (NumLanes == 2) ? MaskVals[i + NumElemsInLane] : -1;
8032 int EltIdx = MaskVals[i];
8034 if ((EltIdx < 0 || EltIdx == (int)i) &&
8035 (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
8038 if (((unsigned)EltIdx == (i + NumElems)) &&
8039 (SndLaneEltIdx < 0 ||
8040 (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
8041 MaskValue |= (1 << i);
8047 *MaskOut = MaskValue;
8051 // Try to lower a shuffle node into a simple blend instruction.
8052 // This function assumes isBlendMask returns true for this
8053 // SuffleVectorSDNode
8054 static SDValue LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
8056 const X86Subtarget *Subtarget,
8057 SelectionDAG &DAG) {
8058 MVT VT = SVOp->getSimpleValueType(0);
8059 MVT EltVT = VT.getVectorElementType();
8060 assert(isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(),
8061 Subtarget->hasInt256() && "Trying to lower a "
8062 "VECTOR_SHUFFLE to a Blend but "
8063 "with the wrong mask"));
8064 SDValue V1 = SVOp->getOperand(0);
8065 SDValue V2 = SVOp->getOperand(1);
8067 unsigned NumElems = VT.getVectorNumElements();
8069 // Convert i32 vectors to floating point if it is not AVX2.
8070 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
8072 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
8073 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
8075 V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
8076 V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
8079 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
8080 DAG.getConstant(MaskValue, MVT::i32));
8081 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
8084 /// In vector type \p VT, return true if the element at index \p InputIdx
8085 /// falls on a different 128-bit lane than \p OutputIdx.
8086 static bool ShuffleCrosses128bitLane(MVT VT, unsigned InputIdx,
8087 unsigned OutputIdx) {
8088 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
8089 return InputIdx * EltSize / 128 != OutputIdx * EltSize / 128;
8092 /// Generate a PSHUFB if possible. Selects elements from \p V1 according to
8093 /// \p MaskVals. MaskVals[OutputIdx] = InputIdx specifies that we want to
8094 /// shuffle the element at InputIdx in V1 to OutputIdx in the result. If \p
8095 /// MaskVals refers to elements outside of \p V1 or is undef (-1), insert a
8097 static SDValue getPSHUFB(ArrayRef<int> MaskVals, SDValue V1, SDLoc &dl,
8098 SelectionDAG &DAG) {
8099 MVT VT = V1.getSimpleValueType();
8100 assert(VT.is128BitVector() || VT.is256BitVector());
8102 MVT EltVT = VT.getVectorElementType();
8103 unsigned EltSizeInBytes = EltVT.getSizeInBits() / 8;
8104 unsigned NumElts = VT.getVectorNumElements();
8106 SmallVector<SDValue, 32> PshufbMask;
8107 for (unsigned OutputIdx = 0; OutputIdx < NumElts; ++OutputIdx) {
8108 int InputIdx = MaskVals[OutputIdx];
8109 unsigned InputByteIdx;
8111 if (InputIdx < 0 || NumElts <= (unsigned)InputIdx)
8112 InputByteIdx = 0x80;
8114 // Cross lane is not allowed.
8115 if (ShuffleCrosses128bitLane(VT, InputIdx, OutputIdx))
8117 InputByteIdx = InputIdx * EltSizeInBytes;
8118 // Index is an byte offset within the 128-bit lane.
8119 InputByteIdx &= 0xf;
8122 for (unsigned j = 0; j < EltSizeInBytes; ++j) {
8123 PshufbMask.push_back(DAG.getConstant(InputByteIdx, MVT::i8));
8124 if (InputByteIdx != 0x80)
8129 MVT ShufVT = MVT::getVectorVT(MVT::i8, PshufbMask.size());
8131 V1 = DAG.getNode(ISD::BITCAST, dl, ShufVT, V1);
8132 return DAG.getNode(X86ISD::PSHUFB, dl, ShufVT, V1,
8133 DAG.getNode(ISD::BUILD_VECTOR, dl, ShufVT, PshufbMask));
8136 // v8i16 shuffles - Prefer shuffles in the following order:
8137 // 1. [all] pshuflw, pshufhw, optional move
8138 // 2. [ssse3] 1 x pshufb
8139 // 3. [ssse3] 2 x pshufb + 1 x por
8140 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
8142 LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
8143 SelectionDAG &DAG) {
8144 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8145 SDValue V1 = SVOp->getOperand(0);
8146 SDValue V2 = SVOp->getOperand(1);
8148 SmallVector<int, 8> MaskVals;
8150 // Determine if more than 1 of the words in each of the low and high quadwords
8151 // of the result come from the same quadword of one of the two inputs. Undef
8152 // mask values count as coming from any quadword, for better codegen.
8154 // Lo/HiQuad[i] = j indicates how many words from the ith quad of the input
8155 // feeds this quad. For i, 0 and 1 refer to V1, 2 and 3 refer to V2.
8156 unsigned LoQuad[] = { 0, 0, 0, 0 };
8157 unsigned HiQuad[] = { 0, 0, 0, 0 };
8158 // Indices of quads used.
8159 std::bitset<4> InputQuads;
8160 for (unsigned i = 0; i < 8; ++i) {
8161 unsigned *Quad = i < 4 ? LoQuad : HiQuad;
8162 int EltIdx = SVOp->getMaskElt(i);
8163 MaskVals.push_back(EltIdx);
8172 InputQuads.set(EltIdx / 4);
8175 int BestLoQuad = -1;
8176 unsigned MaxQuad = 1;
8177 for (unsigned i = 0; i < 4; ++i) {
8178 if (LoQuad[i] > MaxQuad) {
8180 MaxQuad = LoQuad[i];
8184 int BestHiQuad = -1;
8186 for (unsigned i = 0; i < 4; ++i) {
8187 if (HiQuad[i] > MaxQuad) {
8189 MaxQuad = HiQuad[i];
8193 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
8194 // of the two input vectors, shuffle them into one input vector so only a
8195 // single pshufb instruction is necessary. If there are more than 2 input
8196 // quads, disable the next transformation since it does not help SSSE3.
8197 bool V1Used = InputQuads[0] || InputQuads[1];
8198 bool V2Used = InputQuads[2] || InputQuads[3];
8199 if (Subtarget->hasSSSE3()) {
8200 if (InputQuads.count() == 2 && V1Used && V2Used) {
8201 BestLoQuad = InputQuads[0] ? 0 : 1;
8202 BestHiQuad = InputQuads[2] ? 2 : 3;
8204 if (InputQuads.count() > 2) {
8210 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
8211 // the shuffle mask. If a quad is scored as -1, that means that it contains
8212 // words from all 4 input quadwords.
8214 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
8216 BestLoQuad < 0 ? 0 : BestLoQuad,
8217 BestHiQuad < 0 ? 1 : BestHiQuad
8219 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
8220 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
8221 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
8222 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
8224 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
8225 // source words for the shuffle, to aid later transformations.
8226 bool AllWordsInNewV = true;
8227 bool InOrder[2] = { true, true };
8228 for (unsigned i = 0; i != 8; ++i) {
8229 int idx = MaskVals[i];
8231 InOrder[i/4] = false;
8232 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
8234 AllWordsInNewV = false;
8238 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
8239 if (AllWordsInNewV) {
8240 for (int i = 0; i != 8; ++i) {
8241 int idx = MaskVals[i];
8244 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
8245 if ((idx != i) && idx < 4)
8247 if ((idx != i) && idx > 3)
8256 // If we've eliminated the use of V2, and the new mask is a pshuflw or
8257 // pshufhw, that's as cheap as it gets. Return the new shuffle.
8258 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
8259 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
8260 unsigned TargetMask = 0;
8261 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
8262 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
8263 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
8264 TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
8265 getShufflePSHUFLWImmediate(SVOp);
8266 V1 = NewV.getOperand(0);
8267 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
8271 // Promote splats to a larger type which usually leads to more efficient code.
8272 // FIXME: Is this true if pshufb is available?
8273 if (SVOp->isSplat())
8274 return PromoteSplat(SVOp, DAG);
8276 // If we have SSSE3, and all words of the result are from 1 input vector,
8277 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
8278 // is present, fall back to case 4.
8279 if (Subtarget->hasSSSE3()) {
8280 SmallVector<SDValue,16> pshufbMask;
8282 // If we have elements from both input vectors, set the high bit of the
8283 // shuffle mask element to zero out elements that come from V2 in the V1
8284 // mask, and elements that come from V1 in the V2 mask, so that the two
8285 // results can be OR'd together.
8286 bool TwoInputs = V1Used && V2Used;
8287 V1 = getPSHUFB(MaskVals, V1, dl, DAG);
8289 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
8291 // Calculate the shuffle mask for the second input, shuffle it, and
8292 // OR it with the first shuffled input.
8293 CommuteVectorShuffleMask(MaskVals, 8);
8294 V2 = getPSHUFB(MaskVals, V2, dl, DAG);
8295 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
8296 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
8299 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
8300 // and update MaskVals with new element order.
8301 std::bitset<8> InOrder;
8302 if (BestLoQuad >= 0) {
8303 int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
8304 for (int i = 0; i != 4; ++i) {
8305 int idx = MaskVals[i];
8308 } else if ((idx / 4) == BestLoQuad) {
8313 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
8316 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
8317 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
8318 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
8320 getShufflePSHUFLWImmediate(SVOp), DAG);
8324 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
8325 // and update MaskVals with the new element order.
8326 if (BestHiQuad >= 0) {
8327 int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
8328 for (unsigned i = 4; i != 8; ++i) {
8329 int idx = MaskVals[i];
8332 } else if ((idx / 4) == BestHiQuad) {
8333 MaskV[i] = (idx & 3) + 4;
8337 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
8340 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
8341 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
8342 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
8344 getShufflePSHUFHWImmediate(SVOp), DAG);
8348 // In case BestHi & BestLo were both -1, which means each quadword has a word
8349 // from each of the four input quadwords, calculate the InOrder bitvector now
8350 // before falling through to the insert/extract cleanup.
8351 if (BestLoQuad == -1 && BestHiQuad == -1) {
8353 for (int i = 0; i != 8; ++i)
8354 if (MaskVals[i] < 0 || MaskVals[i] == i)
8358 // The other elements are put in the right place using pextrw and pinsrw.
8359 for (unsigned i = 0; i != 8; ++i) {
8362 int EltIdx = MaskVals[i];
8365 SDValue ExtOp = (EltIdx < 8) ?
8366 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
8367 DAG.getIntPtrConstant(EltIdx)) :
8368 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
8369 DAG.getIntPtrConstant(EltIdx - 8));
8370 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
8371 DAG.getIntPtrConstant(i));
8376 /// \brief v16i16 shuffles
8378 /// FIXME: We only support generation of a single pshufb currently. We can
8379 /// generalize the other applicable cases from LowerVECTOR_SHUFFLEv8i16 as
8380 /// well (e.g 2 x pshufb + 1 x por).
8382 LowerVECTOR_SHUFFLEv16i16(SDValue Op, SelectionDAG &DAG) {
8383 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8384 SDValue V1 = SVOp->getOperand(0);
8385 SDValue V2 = SVOp->getOperand(1);
8388 if (V2.getOpcode() != ISD::UNDEF)
8391 SmallVector<int, 16> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
8392 return getPSHUFB(MaskVals, V1, dl, DAG);
8395 // v16i8 shuffles - Prefer shuffles in the following order:
8396 // 1. [ssse3] 1 x pshufb
8397 // 2. [ssse3] 2 x pshufb + 1 x por
8398 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
8399 static SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
8400 const X86Subtarget* Subtarget,
8401 SelectionDAG &DAG) {
8402 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8403 SDValue V1 = SVOp->getOperand(0);
8404 SDValue V2 = SVOp->getOperand(1);
8406 ArrayRef<int> MaskVals = SVOp->getMask();
8408 // Promote splats to a larger type which usually leads to more efficient code.
8409 // FIXME: Is this true if pshufb is available?
8410 if (SVOp->isSplat())
8411 return PromoteSplat(SVOp, DAG);
8413 // If we have SSSE3, case 1 is generated when all result bytes come from
8414 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
8415 // present, fall back to case 3.
8417 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
8418 if (Subtarget->hasSSSE3()) {
8419 SmallVector<SDValue,16> pshufbMask;
8421 // If all result elements are from one input vector, then only translate
8422 // undef mask values to 0x80 (zero out result) in the pshufb mask.
8424 // Otherwise, we have elements from both input vectors, and must zero out
8425 // elements that come from V2 in the first mask, and V1 in the second mask
8426 // so that we can OR them together.
8427 for (unsigned i = 0; i != 16; ++i) {
8428 int EltIdx = MaskVals[i];
8429 if (EltIdx < 0 || EltIdx >= 16)
8431 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
8433 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
8434 DAG.getNode(ISD::BUILD_VECTOR, dl,
8435 MVT::v16i8, pshufbMask));
8437 // As PSHUFB will zero elements with negative indices, it's safe to ignore
8438 // the 2nd operand if it's undefined or zero.
8439 if (V2.getOpcode() == ISD::UNDEF ||
8440 ISD::isBuildVectorAllZeros(V2.getNode()))
8443 // Calculate the shuffle mask for the second input, shuffle it, and
8444 // OR it with the first shuffled input.
8446 for (unsigned i = 0; i != 16; ++i) {
8447 int EltIdx = MaskVals[i];
8448 EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
8449 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
8451 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
8452 DAG.getNode(ISD::BUILD_VECTOR, dl,
8453 MVT::v16i8, pshufbMask));
8454 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
8457 // No SSSE3 - Calculate in place words and then fix all out of place words
8458 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
8459 // the 16 different words that comprise the two doublequadword input vectors.
8460 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
8461 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
8463 for (int i = 0; i != 8; ++i) {
8464 int Elt0 = MaskVals[i*2];
8465 int Elt1 = MaskVals[i*2+1];
8467 // This word of the result is all undef, skip it.
8468 if (Elt0 < 0 && Elt1 < 0)
8471 // This word of the result is already in the correct place, skip it.
8472 if ((Elt0 == i*2) && (Elt1 == i*2+1))
8475 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
8476 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
8479 // If Elt0 and Elt1 are defined, are consecutive, and can be load
8480 // using a single extract together, load it and store it.
8481 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
8482 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
8483 DAG.getIntPtrConstant(Elt1 / 2));
8484 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
8485 DAG.getIntPtrConstant(i));
8489 // If Elt1 is defined, extract it from the appropriate source. If the
8490 // source byte is not also odd, shift the extracted word left 8 bits
8491 // otherwise clear the bottom 8 bits if we need to do an or.
8493 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
8494 DAG.getIntPtrConstant(Elt1 / 2));
8495 if ((Elt1 & 1) == 0)
8496 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
8498 TLI.getShiftAmountTy(InsElt.getValueType())));
8500 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
8501 DAG.getConstant(0xFF00, MVT::i16));
8503 // If Elt0 is defined, extract it from the appropriate source. If the
8504 // source byte is not also even, shift the extracted word right 8 bits. If
8505 // Elt1 was also defined, OR the extracted values together before
8506 // inserting them in the result.
8508 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
8509 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
8510 if ((Elt0 & 1) != 0)
8511 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
8513 TLI.getShiftAmountTy(InsElt0.getValueType())));
8515 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
8516 DAG.getConstant(0x00FF, MVT::i16));
8517 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
8520 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
8521 DAG.getIntPtrConstant(i));
8523 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
8526 // v32i8 shuffles - Translate to VPSHUFB if possible.
8528 SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
8529 const X86Subtarget *Subtarget,
8530 SelectionDAG &DAG) {
8531 MVT VT = SVOp->getSimpleValueType(0);
8532 SDValue V1 = SVOp->getOperand(0);
8533 SDValue V2 = SVOp->getOperand(1);
8535 SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
8537 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
8538 bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
8539 bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
8541 // VPSHUFB may be generated if
8542 // (1) one of input vector is undefined or zeroinitializer.
8543 // The mask value 0x80 puts 0 in the corresponding slot of the vector.
8544 // And (2) the mask indexes don't cross the 128-bit lane.
8545 if (VT != MVT::v32i8 || !Subtarget->hasInt256() ||
8546 (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
8549 if (V1IsAllZero && !V2IsAllZero) {
8550 CommuteVectorShuffleMask(MaskVals, 32);
8553 return getPSHUFB(MaskVals, V1, dl, DAG);
8556 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
8557 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
8558 /// done when every pair / quad of shuffle mask elements point to elements in
8559 /// the right sequence. e.g.
8560 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
8562 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
8563 SelectionDAG &DAG) {
8564 MVT VT = SVOp->getSimpleValueType(0);
8566 unsigned NumElems = VT.getVectorNumElements();
8569 switch (VT.SimpleTy) {
8570 default: llvm_unreachable("Unexpected!");
8573 return SDValue(SVOp, 0);
8574 case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
8575 case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
8576 case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
8577 case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
8578 case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
8579 case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
8582 SmallVector<int, 8> MaskVec;
8583 for (unsigned i = 0; i != NumElems; i += Scale) {
8585 for (unsigned j = 0; j != Scale; ++j) {
8586 int EltIdx = SVOp->getMaskElt(i+j);
8590 StartIdx = (EltIdx / Scale);
8591 if (EltIdx != (int)(StartIdx*Scale + j))
8594 MaskVec.push_back(StartIdx);
8597 SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
8598 SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
8599 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
8602 /// getVZextMovL - Return a zero-extending vector move low node.
8604 static SDValue getVZextMovL(MVT VT, MVT OpVT,
8605 SDValue SrcOp, SelectionDAG &DAG,
8606 const X86Subtarget *Subtarget, SDLoc dl) {
8607 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
8608 LoadSDNode *LD = nullptr;
8609 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
8610 LD = dyn_cast<LoadSDNode>(SrcOp);
8612 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
8614 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
8615 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
8616 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
8617 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
8618 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
8620 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
8621 return DAG.getNode(ISD::BITCAST, dl, VT,
8622 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
8623 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
8631 return DAG.getNode(ISD::BITCAST, dl, VT,
8632 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
8633 DAG.getNode(ISD::BITCAST, dl,
8637 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
8638 /// which could not be matched by any known target speficic shuffle
8640 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
8642 SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
8643 if (NewOp.getNode())
8646 MVT VT = SVOp->getSimpleValueType(0);
8648 unsigned NumElems = VT.getVectorNumElements();
8649 unsigned NumLaneElems = NumElems / 2;
8652 MVT EltVT = VT.getVectorElementType();
8653 MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
8656 SmallVector<int, 16> Mask;
8657 for (unsigned l = 0; l < 2; ++l) {
8658 // Build a shuffle mask for the output, discovering on the fly which
8659 // input vectors to use as shuffle operands (recorded in InputUsed).
8660 // If building a suitable shuffle vector proves too hard, then bail
8661 // out with UseBuildVector set.
8662 bool UseBuildVector = false;
8663 int InputUsed[2] = { -1, -1 }; // Not yet discovered.
8664 unsigned LaneStart = l * NumLaneElems;
8665 for (unsigned i = 0; i != NumLaneElems; ++i) {
8666 // The mask element. This indexes into the input.
8667 int Idx = SVOp->getMaskElt(i+LaneStart);
8669 // the mask element does not index into any input vector.
8674 // The input vector this mask element indexes into.
8675 int Input = Idx / NumLaneElems;
8677 // Turn the index into an offset from the start of the input vector.
8678 Idx -= Input * NumLaneElems;
8680 // Find or create a shuffle vector operand to hold this input.
8682 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
8683 if (InputUsed[OpNo] == Input)
8684 // This input vector is already an operand.
8686 if (InputUsed[OpNo] < 0) {
8687 // Create a new operand for this input vector.
8688 InputUsed[OpNo] = Input;
8693 if (OpNo >= array_lengthof(InputUsed)) {
8694 // More than two input vectors used! Give up on trying to create a
8695 // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
8696 UseBuildVector = true;
8700 // Add the mask index for the new shuffle vector.
8701 Mask.push_back(Idx + OpNo * NumLaneElems);
8704 if (UseBuildVector) {
8705 SmallVector<SDValue, 16> SVOps;
8706 for (unsigned i = 0; i != NumLaneElems; ++i) {
8707 // The mask element. This indexes into the input.
8708 int Idx = SVOp->getMaskElt(i+LaneStart);
8710 SVOps.push_back(DAG.getUNDEF(EltVT));
8714 // The input vector this mask element indexes into.
8715 int Input = Idx / NumElems;
8717 // Turn the index into an offset from the start of the input vector.
8718 Idx -= Input * NumElems;
8720 // Extract the vector element by hand.
8721 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
8722 SVOp->getOperand(Input),
8723 DAG.getIntPtrConstant(Idx)));
8726 // Construct the output using a BUILD_VECTOR.
8727 Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, SVOps);
8728 } else if (InputUsed[0] < 0) {
8729 // No input vectors were used! The result is undefined.
8730 Output[l] = DAG.getUNDEF(NVT);
8732 SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
8733 (InputUsed[0] % 2) * NumLaneElems,
8735 // If only one input was used, use an undefined vector for the other.
8736 SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
8737 Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
8738 (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
8739 // At least one input vector was used. Create a new shuffle vector.
8740 Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
8746 // Concatenate the result back
8747 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
8750 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
8751 /// 4 elements, and match them with several different shuffle types.
8753 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
8754 SDValue V1 = SVOp->getOperand(0);
8755 SDValue V2 = SVOp->getOperand(1);
8757 MVT VT = SVOp->getSimpleValueType(0);
8759 assert(VT.is128BitVector() && "Unsupported vector size");
8761 std::pair<int, int> Locs[4];
8762 int Mask1[] = { -1, -1, -1, -1 };
8763 SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
8767 for (unsigned i = 0; i != 4; ++i) {
8768 int Idx = PermMask[i];
8770 Locs[i] = std::make_pair(-1, -1);
8772 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
8774 Locs[i] = std::make_pair(0, NumLo);
8778 Locs[i] = std::make_pair(1, NumHi);
8780 Mask1[2+NumHi] = Idx;
8786 if (NumLo <= 2 && NumHi <= 2) {
8787 // If no more than two elements come from either vector. This can be
8788 // implemented with two shuffles. First shuffle gather the elements.
8789 // The second shuffle, which takes the first shuffle as both of its
8790 // vector operands, put the elements into the right order.
8791 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
8793 int Mask2[] = { -1, -1, -1, -1 };
8795 for (unsigned i = 0; i != 4; ++i)
8796 if (Locs[i].first != -1) {
8797 unsigned Idx = (i < 2) ? 0 : 4;
8798 Idx += Locs[i].first * 2 + Locs[i].second;
8802 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
8805 if (NumLo == 3 || NumHi == 3) {
8806 // Otherwise, we must have three elements from one vector, call it X, and
8807 // one element from the other, call it Y. First, use a shufps to build an
8808 // intermediate vector with the one element from Y and the element from X
8809 // that will be in the same half in the final destination (the indexes don't
8810 // matter). Then, use a shufps to build the final vector, taking the half
8811 // containing the element from Y from the intermediate, and the other half
8814 // Normalize it so the 3 elements come from V1.
8815 CommuteVectorShuffleMask(PermMask, 4);
8819 // Find the element from V2.
8821 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
8822 int Val = PermMask[HiIndex];
8829 Mask1[0] = PermMask[HiIndex];
8831 Mask1[2] = PermMask[HiIndex^1];
8833 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
8836 Mask1[0] = PermMask[0];
8837 Mask1[1] = PermMask[1];
8838 Mask1[2] = HiIndex & 1 ? 6 : 4;
8839 Mask1[3] = HiIndex & 1 ? 4 : 6;
8840 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
8843 Mask1[0] = HiIndex & 1 ? 2 : 0;
8844 Mask1[1] = HiIndex & 1 ? 0 : 2;
8845 Mask1[2] = PermMask[2];
8846 Mask1[3] = PermMask[3];
8851 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
8854 // Break it into (shuffle shuffle_hi, shuffle_lo).
8855 int LoMask[] = { -1, -1, -1, -1 };
8856 int HiMask[] = { -1, -1, -1, -1 };
8858 int *MaskPtr = LoMask;
8859 unsigned MaskIdx = 0;
8862 for (unsigned i = 0; i != 4; ++i) {
8869 int Idx = PermMask[i];
8871 Locs[i] = std::make_pair(-1, -1);
8872 } else if (Idx < 4) {
8873 Locs[i] = std::make_pair(MaskIdx, LoIdx);
8874 MaskPtr[LoIdx] = Idx;
8877 Locs[i] = std::make_pair(MaskIdx, HiIdx);
8878 MaskPtr[HiIdx] = Idx;
8883 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
8884 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
8885 int MaskOps[] = { -1, -1, -1, -1 };
8886 for (unsigned i = 0; i != 4; ++i)
8887 if (Locs[i].first != -1)
8888 MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
8889 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
8892 static bool MayFoldVectorLoad(SDValue V) {
8893 while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
8894 V = V.getOperand(0);
8896 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
8897 V = V.getOperand(0);
8898 if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
8899 V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
8900 // BUILD_VECTOR (load), undef
8901 V = V.getOperand(0);
8903 return MayFoldLoad(V);
8907 SDValue getMOVDDup(SDValue &Op, SDLoc &dl, SDValue V1, SelectionDAG &DAG) {
8908 MVT VT = Op.getSimpleValueType();
8910 // Canonizalize to v2f64.
8911 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
8912 return DAG.getNode(ISD::BITCAST, dl, VT,
8913 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
8918 SDValue getMOVLowToHigh(SDValue &Op, SDLoc &dl, SelectionDAG &DAG,
8920 SDValue V1 = Op.getOperand(0);
8921 SDValue V2 = Op.getOperand(1);
8922 MVT VT = Op.getSimpleValueType();
8924 assert(VT != MVT::v2i64 && "unsupported shuffle type");
8926 if (HasSSE2 && VT == MVT::v2f64)
8927 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
8929 // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
8930 return DAG.getNode(ISD::BITCAST, dl, VT,
8931 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
8932 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
8933 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
8937 SDValue getMOVHighToLow(SDValue &Op, SDLoc &dl, SelectionDAG &DAG) {
8938 SDValue V1 = Op.getOperand(0);
8939 SDValue V2 = Op.getOperand(1);
8940 MVT VT = Op.getSimpleValueType();
8942 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
8943 "unsupported shuffle type");
8945 if (V2.getOpcode() == ISD::UNDEF)
8949 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
8953 SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
8954 SDValue V1 = Op.getOperand(0);
8955 SDValue V2 = Op.getOperand(1);
8956 MVT VT = Op.getSimpleValueType();
8957 unsigned NumElems = VT.getVectorNumElements();
8959 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
8960 // operand of these instructions is only memory, so check if there's a
8961 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
8963 bool CanFoldLoad = false;
8965 // Trivial case, when V2 comes from a load.
8966 if (MayFoldVectorLoad(V2))
8969 // When V1 is a load, it can be folded later into a store in isel, example:
8970 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
8972 // (MOVLPSmr addr:$src1, VR128:$src2)
8973 // So, recognize this potential and also use MOVLPS or MOVLPD
8974 else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
8977 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8979 if (HasSSE2 && NumElems == 2)
8980 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
8983 // If we don't care about the second element, proceed to use movss.
8984 if (SVOp->getMaskElt(1) != -1)
8985 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
8988 // movl and movlp will both match v2i64, but v2i64 is never matched by
8989 // movl earlier because we make it strict to avoid messing with the movlp load
8990 // folding logic (see the code above getMOVLP call). Match it here then,
8991 // this is horrible, but will stay like this until we move all shuffle
8992 // matching to x86 specific nodes. Note that for the 1st condition all
8993 // types are matched with movsd.
8995 // FIXME: isMOVLMask should be checked and matched before getMOVLP,
8996 // as to remove this logic from here, as much as possible
8997 if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
8998 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
8999 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
9002 assert(VT != MVT::v4i32 && "unsupported shuffle type");
9004 // Invert the operand order and use SHUFPS to match it.
9005 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
9006 getShuffleSHUFImmediate(SVOp), DAG);
9009 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
9010 SelectionDAG &DAG) {
9012 MVT VT = Load->getSimpleValueType(0);
9013 MVT EVT = VT.getVectorElementType();
9014 SDValue Addr = Load->getOperand(1);
9015 SDValue NewAddr = DAG.getNode(
9016 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
9017 DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
9020 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
9021 DAG.getMachineFunction().getMachineMemOperand(
9022 Load->getMemOperand(), 0, EVT.getStoreSize()));
9026 // It is only safe to call this function if isINSERTPSMask is true for
9027 // this shufflevector mask.
9028 static SDValue getINSERTPS(ShuffleVectorSDNode *SVOp, SDLoc &dl,
9029 SelectionDAG &DAG) {
9030 // Generate an insertps instruction when inserting an f32 from memory onto a
9031 // v4f32 or when copying a member from one v4f32 to another.
9032 // We also use it for transferring i32 from one register to another,
9033 // since it simply copies the same bits.
9034 // If we're transferring an i32 from memory to a specific element in a
9035 // register, we output a generic DAG that will match the PINSRD
9037 MVT VT = SVOp->getSimpleValueType(0);
9038 MVT EVT = VT.getVectorElementType();
9039 SDValue V1 = SVOp->getOperand(0);
9040 SDValue V2 = SVOp->getOperand(1);
9041 auto Mask = SVOp->getMask();
9042 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
9043 "unsupported vector type for insertps/pinsrd");
9045 auto FromV1Predicate = [](const int &i) { return i < 4 && i > -1; };
9046 auto FromV2Predicate = [](const int &i) { return i >= 4; };
9047 int FromV1 = std::count_if(Mask.begin(), Mask.end(), FromV1Predicate);
9055 DestIndex = std::find_if(Mask.begin(), Mask.end(), FromV1Predicate) -
9058 // If we have 1 element from each vector, we have to check if we're
9059 // changing V1's element's place. If so, we're done. Otherwise, we
9060 // should assume we're changing V2's element's place and behave
9062 int FromV2 = std::count_if(Mask.begin(), Mask.end(), FromV2Predicate);
9063 if (FromV1 == FromV2 && DestIndex == Mask[DestIndex] % 4) {
9067 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
9070 assert(std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 &&
9071 "More than one element from V1 and from V2, or no elements from one "
9072 "of the vectors. This case should not have returned true from "
9077 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
9080 // Get an index into the source vector in the range [0,4) (the mask is
9081 // in the range [0,8) because it can address V1 and V2)
9082 unsigned SrcIndex = Mask[DestIndex] % 4;
9083 if (MayFoldLoad(From)) {
9084 // Trivial case, when From comes from a load and is only used by the
9085 // shuffle. Make it use insertps from the vector that we need from that
9088 NarrowVectorLoadToElement(cast<LoadSDNode>(From), SrcIndex, DAG);
9089 if (!NewLoad.getNode())
9092 if (EVT == MVT::f32) {
9093 // Create this as a scalar to vector to match the instruction pattern.
9094 SDValue LoadScalarToVector =
9095 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, NewLoad);
9096 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4);
9097 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, LoadScalarToVector,
9099 } else { // EVT == MVT::i32
9100 // If we're getting an i32 from memory, use an INSERT_VECTOR_ELT
9101 // instruction, to match the PINSRD instruction, which loads an i32 to a
9102 // certain vector element.
9103 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, To, NewLoad,
9104 DAG.getConstant(DestIndex, MVT::i32));
9108 // Vector-element-to-vector
9109 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4 | SrcIndex << 6);
9110 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, From, InsertpsMask);
9113 // Reduce a vector shuffle to zext.
9114 static SDValue LowerVectorIntExtend(SDValue Op, const X86Subtarget *Subtarget,
9115 SelectionDAG &DAG) {
9116 // PMOVZX is only available from SSE41.
9117 if (!Subtarget->hasSSE41())
9120 MVT VT = Op.getSimpleValueType();
9122 // Only AVX2 support 256-bit vector integer extending.
9123 if (!Subtarget->hasInt256() && VT.is256BitVector())
9126 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9128 SDValue V1 = Op.getOperand(0);
9129 SDValue V2 = Op.getOperand(1);
9130 unsigned NumElems = VT.getVectorNumElements();
9132 // Extending is an unary operation and the element type of the source vector
9133 // won't be equal to or larger than i64.
9134 if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
9135 VT.getVectorElementType() == MVT::i64)
9138 // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
9139 unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
9140 while ((1U << Shift) < NumElems) {
9141 if (SVOp->getMaskElt(1U << Shift) == 1)
9144 // The maximal ratio is 8, i.e. from i8 to i64.
9149 // Check the shuffle mask.
9150 unsigned Mask = (1U << Shift) - 1;
9151 for (unsigned i = 0; i != NumElems; ++i) {
9152 int EltIdx = SVOp->getMaskElt(i);
9153 if ((i & Mask) != 0 && EltIdx != -1)
9155 if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
9159 unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
9160 MVT NeVT = MVT::getIntegerVT(NBits);
9161 MVT NVT = MVT::getVectorVT(NeVT, NumElems >> Shift);
9163 if (!DAG.getTargetLoweringInfo().isTypeLegal(NVT))
9166 // Simplify the operand as it's prepared to be fed into shuffle.
9167 unsigned SignificantBits = NVT.getSizeInBits() >> Shift;
9168 if (V1.getOpcode() == ISD::BITCAST &&
9169 V1.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
9170 V1.getOperand(0).getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
9171 V1.getOperand(0).getOperand(0)
9172 .getSimpleValueType().getSizeInBits() == SignificantBits) {
9173 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
9174 SDValue V = V1.getOperand(0).getOperand(0).getOperand(0);
9175 ConstantSDNode *CIdx =
9176 dyn_cast<ConstantSDNode>(V1.getOperand(0).getOperand(0).getOperand(1));
9177 // If it's foldable, i.e. normal load with single use, we will let code
9178 // selection to fold it. Otherwise, we will short the conversion sequence.
9179 if (CIdx && CIdx->getZExtValue() == 0 &&
9180 (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse())) {
9181 MVT FullVT = V.getSimpleValueType();
9182 MVT V1VT = V1.getSimpleValueType();
9183 if (FullVT.getSizeInBits() > V1VT.getSizeInBits()) {
9184 // The "ext_vec_elt" node is wider than the result node.
9185 // In this case we should extract subvector from V.
9186 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast (extract_subvector x)).
9187 unsigned Ratio = FullVT.getSizeInBits() / V1VT.getSizeInBits();
9188 MVT SubVecVT = MVT::getVectorVT(FullVT.getVectorElementType(),
9189 FullVT.getVectorNumElements()/Ratio);
9190 V = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, V,
9191 DAG.getIntPtrConstant(0));
9193 V1 = DAG.getNode(ISD::BITCAST, DL, V1VT, V);
9197 return DAG.getNode(ISD::BITCAST, DL, VT,
9198 DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
9201 static SDValue NormalizeVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
9202 SelectionDAG &DAG) {
9203 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9204 MVT VT = Op.getSimpleValueType();
9206 SDValue V1 = Op.getOperand(0);
9207 SDValue V2 = Op.getOperand(1);
9209 if (isZeroShuffle(SVOp))
9210 return getZeroVector(VT, Subtarget, DAG, dl);
9212 // Handle splat operations
9213 if (SVOp->isSplat()) {
9214 // Use vbroadcast whenever the splat comes from a foldable load
9215 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
9216 if (Broadcast.getNode())
9220 // Check integer expanding shuffles.
9221 SDValue NewOp = LowerVectorIntExtend(Op, Subtarget, DAG);
9222 if (NewOp.getNode())
9225 // If the shuffle can be profitably rewritten as a narrower shuffle, then
9227 if (VT == MVT::v8i16 || VT == MVT::v16i8 || VT == MVT::v16i16 ||
9229 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
9230 if (NewOp.getNode())
9231 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
9232 } else if (VT.is128BitVector() && Subtarget->hasSSE2()) {
9233 // FIXME: Figure out a cleaner way to do this.
9234 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
9235 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
9236 if (NewOp.getNode()) {
9237 MVT NewVT = NewOp.getSimpleValueType();
9238 if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
9239 NewVT, true, false))
9240 return getVZextMovL(VT, NewVT, NewOp.getOperand(0), DAG, Subtarget,
9243 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
9244 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
9245 if (NewOp.getNode()) {
9246 MVT NewVT = NewOp.getSimpleValueType();
9247 if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
9248 return getVZextMovL(VT, NewVT, NewOp.getOperand(1), DAG, Subtarget,
9257 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
9258 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9259 SDValue V1 = Op.getOperand(0);
9260 SDValue V2 = Op.getOperand(1);
9261 MVT VT = Op.getSimpleValueType();
9263 unsigned NumElems = VT.getVectorNumElements();
9264 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
9265 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
9266 bool V1IsSplat = false;
9267 bool V2IsSplat = false;
9268 bool HasSSE2 = Subtarget->hasSSE2();
9269 bool HasFp256 = Subtarget->hasFp256();
9270 bool HasInt256 = Subtarget->hasInt256();
9271 MachineFunction &MF = DAG.getMachineFunction();
9272 bool OptForSize = MF.getFunction()->getAttributes().
9273 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
9275 // Check if we should use the experimental vector shuffle lowering. If so,
9276 // delegate completely to that code path.
9277 if (ExperimentalVectorShuffleLowering)
9278 return lowerVectorShuffle(Op, Subtarget, DAG);
9280 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
9282 if (V1IsUndef && V2IsUndef)
9283 return DAG.getUNDEF(VT);
9285 // When we create a shuffle node we put the UNDEF node to second operand,
9286 // but in some cases the first operand may be transformed to UNDEF.
9287 // In this case we should just commute the node.
9289 return DAG.getCommutedVectorShuffle(*SVOp);
9291 // Vector shuffle lowering takes 3 steps:
9293 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
9294 // narrowing and commutation of operands should be handled.
9295 // 2) Matching of shuffles with known shuffle masks to x86 target specific
9297 // 3) Rewriting of unmatched masks into new generic shuffle operations,
9298 // so the shuffle can be broken into other shuffles and the legalizer can
9299 // try the lowering again.
9301 // The general idea is that no vector_shuffle operation should be left to
9302 // be matched during isel, all of them must be converted to a target specific
9305 // Normalize the input vectors. Here splats, zeroed vectors, profitable
9306 // narrowing and commutation of operands should be handled. The actual code
9307 // doesn't include all of those, work in progress...
9308 SDValue NewOp = NormalizeVectorShuffle(Op, Subtarget, DAG);
9309 if (NewOp.getNode())
9312 SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
9314 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
9315 // unpckh_undef). Only use pshufd if speed is more important than size.
9316 if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256))
9317 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
9318 if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256))
9319 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
9321 if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
9322 V2IsUndef && MayFoldVectorLoad(V1))
9323 return getMOVDDup(Op, dl, V1, DAG);
9325 if (isMOVHLPS_v_undef_Mask(M, VT))
9326 return getMOVHighToLow(Op, dl, DAG);
9328 // Use to match splats
9329 if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef &&
9330 (VT == MVT::v2f64 || VT == MVT::v2i64))
9331 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
9333 if (isPSHUFDMask(M, VT)) {
9334 // The actual implementation will match the mask in the if above and then
9335 // during isel it can match several different instructions, not only pshufd
9336 // as its name says, sad but true, emulate the behavior for now...
9337 if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
9338 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
9340 unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
9342 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
9343 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
9345 if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
9346 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1, TargetMask,
9349 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
9353 if (isPALIGNRMask(M, VT, Subtarget))
9354 return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
9355 getShufflePALIGNRImmediate(SVOp),
9358 // Check if this can be converted into a logical shift.
9359 bool isLeft = false;
9362 bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
9363 if (isShift && ShVal.hasOneUse()) {
9364 // If the shifted value has multiple uses, it may be cheaper to use
9365 // v_set0 + movlhps or movhlps, etc.
9366 MVT EltVT = VT.getVectorElementType();
9367 ShAmt *= EltVT.getSizeInBits();
9368 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
9371 if (isMOVLMask(M, VT)) {
9372 if (ISD::isBuildVectorAllZeros(V1.getNode()))
9373 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
9374 if (!isMOVLPMask(M, VT)) {
9375 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
9376 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
9378 if (VT == MVT::v4i32 || VT == MVT::v4f32)
9379 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
9383 // FIXME: fold these into legal mask.
9384 if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256))
9385 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
9387 if (isMOVHLPSMask(M, VT))
9388 return getMOVHighToLow(Op, dl, DAG);
9390 if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
9391 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
9393 if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
9394 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
9396 if (isMOVLPMask(M, VT))
9397 return getMOVLP(Op, dl, DAG, HasSSE2);
9399 if (ShouldXformToMOVHLPS(M, VT) ||
9400 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
9401 return DAG.getCommutedVectorShuffle(*SVOp);
9404 // No better options. Use a vshldq / vsrldq.
9405 MVT EltVT = VT.getVectorElementType();
9406 ShAmt *= EltVT.getSizeInBits();
9407 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
9410 bool Commuted = false;
9411 // FIXME: This should also accept a bitcast of a splat? Be careful, not
9412 // 1,1,1,1 -> v8i16 though.
9413 BitVector UndefElements;
9414 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V1.getNode()))
9415 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
9417 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V2.getNode()))
9418 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
9421 // Canonicalize the splat or undef, if present, to be on the RHS.
9422 if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
9423 CommuteVectorShuffleMask(M, NumElems);
9425 std::swap(V1IsSplat, V2IsSplat);
9429 if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
9430 // Shuffling low element of v1 into undef, just return v1.
9433 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
9434 // the instruction selector will not match, so get a canonical MOVL with
9435 // swapped operands to undo the commute.
9436 return getMOVL(DAG, dl, VT, V2, V1);
9439 if (isUNPCKLMask(M, VT, HasInt256))
9440 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
9442 if (isUNPCKHMask(M, VT, HasInt256))
9443 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
9446 // Normalize mask so all entries that point to V2 points to its first
9447 // element then try to match unpck{h|l} again. If match, return a
9448 // new vector_shuffle with the corrected mask.p
9449 SmallVector<int, 8> NewMask(M.begin(), M.end());
9450 NormalizeMask(NewMask, NumElems);
9451 if (isUNPCKLMask(NewMask, VT, HasInt256, true))
9452 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
9453 if (isUNPCKHMask(NewMask, VT, HasInt256, true))
9454 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
9458 // Commute is back and try unpck* again.
9459 // FIXME: this seems wrong.
9460 CommuteVectorShuffleMask(M, NumElems);
9462 std::swap(V1IsSplat, V2IsSplat);
9464 if (isUNPCKLMask(M, VT, HasInt256))
9465 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
9467 if (isUNPCKHMask(M, VT, HasInt256))
9468 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
9471 // Normalize the node to match x86 shuffle ops if needed
9472 if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true)))
9473 return DAG.getCommutedVectorShuffle(*SVOp);
9475 // The checks below are all present in isShuffleMaskLegal, but they are
9476 // inlined here right now to enable us to directly emit target specific
9477 // nodes, and remove one by one until they don't return Op anymore.
9479 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
9480 SVOp->getSplatIndex() == 0 && V2IsUndef) {
9481 if (VT == MVT::v2f64 || VT == MVT::v2i64)
9482 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
9485 if (isPSHUFHWMask(M, VT, HasInt256))
9486 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
9487 getShufflePSHUFHWImmediate(SVOp),
9490 if (isPSHUFLWMask(M, VT, HasInt256))
9491 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
9492 getShufflePSHUFLWImmediate(SVOp),
9496 if (isBlendMask(M, VT, Subtarget->hasSSE41(), Subtarget->hasInt256(),
9498 return LowerVECTOR_SHUFFLEtoBlend(SVOp, MaskValue, Subtarget, DAG);
9500 if (isSHUFPMask(M, VT))
9501 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
9502 getShuffleSHUFImmediate(SVOp), DAG);
9504 if (isUNPCKL_v_undef_Mask(M, VT, HasInt256))
9505 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
9506 if (isUNPCKH_v_undef_Mask(M, VT, HasInt256))
9507 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
9509 //===--------------------------------------------------------------------===//
9510 // Generate target specific nodes for 128 or 256-bit shuffles only
9511 // supported in the AVX instruction set.
9514 // Handle VMOVDDUPY permutations
9515 if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256))
9516 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
9518 // Handle VPERMILPS/D* permutations
9519 if (isVPERMILPMask(M, VT)) {
9520 if ((HasInt256 && VT == MVT::v8i32) || VT == MVT::v16i32)
9521 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
9522 getShuffleSHUFImmediate(SVOp), DAG);
9523 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1,
9524 getShuffleSHUFImmediate(SVOp), DAG);
9528 if (VT.is512BitVector() && isINSERT64x4Mask(M, VT, &Idx))
9529 return Insert256BitVector(V1, Extract256BitVector(V2, 0, DAG, dl),
9530 Idx*(NumElems/2), DAG, dl);
9532 // Handle VPERM2F128/VPERM2I128 permutations
9533 if (isVPERM2X128Mask(M, VT, HasFp256))
9534 return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
9535 V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
9537 if (Subtarget->hasSSE41() && isINSERTPSMask(M, VT))
9538 return getINSERTPS(SVOp, dl, DAG);
9541 if (V2IsUndef && HasInt256 && isPermImmMask(M, VT, Imm8))
9542 return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, Imm8, DAG);
9544 if ((V2IsUndef && HasInt256 && VT.is256BitVector() && NumElems == 8) ||
9545 VT.is512BitVector()) {
9546 MVT MaskEltVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits());
9547 MVT MaskVectorVT = MVT::getVectorVT(MaskEltVT, NumElems);
9548 SmallVector<SDValue, 16> permclMask;
9549 for (unsigned i = 0; i != NumElems; ++i) {
9550 permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MaskEltVT));
9553 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVectorVT, permclMask);
9555 // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
9556 return DAG.getNode(X86ISD::VPERMV, dl, VT,
9557 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
9558 return DAG.getNode(X86ISD::VPERMV3, dl, VT, V1,
9559 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V2);
9562 //===--------------------------------------------------------------------===//
9563 // Since no target specific shuffle was selected for this generic one,
9564 // lower it into other known shuffles. FIXME: this isn't true yet, but
9565 // this is the plan.
9568 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
9569 if (VT == MVT::v8i16) {
9570 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
9571 if (NewOp.getNode())
9575 if (VT == MVT::v16i16 && Subtarget->hasInt256()) {
9576 SDValue NewOp = LowerVECTOR_SHUFFLEv16i16(Op, DAG);
9577 if (NewOp.getNode())
9581 if (VT == MVT::v16i8) {
9582 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG);
9583 if (NewOp.getNode())
9587 if (VT == MVT::v32i8) {
9588 SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
9589 if (NewOp.getNode())
9593 // Handle all 128-bit wide vectors with 4 elements, and match them with
9594 // several different shuffle types.
9595 if (NumElems == 4 && VT.is128BitVector())
9596 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
9598 // Handle general 256-bit shuffles
9599 if (VT.is256BitVector())
9600 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
9605 // This function assumes its argument is a BUILD_VECTOR of constants or
9606 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
9608 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
9609 unsigned &MaskValue) {
9611 unsigned NumElems = BuildVector->getNumOperands();
9612 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
9613 unsigned NumLanes = (NumElems - 1) / 8 + 1;
9614 unsigned NumElemsInLane = NumElems / NumLanes;
9616 // Blend for v16i16 should be symetric for the both lanes.
9617 for (unsigned i = 0; i < NumElemsInLane; ++i) {
9618 SDValue EltCond = BuildVector->getOperand(i);
9619 SDValue SndLaneEltCond =
9620 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
9622 int Lane1Cond = -1, Lane2Cond = -1;
9623 if (isa<ConstantSDNode>(EltCond))
9624 Lane1Cond = !isZero(EltCond);
9625 if (isa<ConstantSDNode>(SndLaneEltCond))
9626 Lane2Cond = !isZero(SndLaneEltCond);
9628 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
9629 // Lane1Cond != 0, means we want the first argument.
9630 // Lane1Cond == 0, means we want the second argument.
9631 // The encoding of this argument is 0 for the first argument, 1
9632 // for the second. Therefore, invert the condition.
9633 MaskValue |= !Lane1Cond << i;
9634 else if (Lane1Cond < 0)
9635 MaskValue |= !Lane2Cond << i;
9642 // Try to lower a vselect node into a simple blend instruction.
9643 static SDValue LowerVSELECTtoBlend(SDValue Op, const X86Subtarget *Subtarget,
9644 SelectionDAG &DAG) {
9645 SDValue Cond = Op.getOperand(0);
9646 SDValue LHS = Op.getOperand(1);
9647 SDValue RHS = Op.getOperand(2);
9649 MVT VT = Op.getSimpleValueType();
9650 MVT EltVT = VT.getVectorElementType();
9651 unsigned NumElems = VT.getVectorNumElements();
9653 // There is no blend with immediate in AVX-512.
9654 if (VT.is512BitVector())
9657 if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
9659 if (!Subtarget->hasInt256() && VT == MVT::v16i16)
9662 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
9665 // Check the mask for BLEND and build the value.
9666 unsigned MaskValue = 0;
9667 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
9670 // Convert i32 vectors to floating point if it is not AVX2.
9671 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
9673 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
9674 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
9676 LHS = DAG.getNode(ISD::BITCAST, dl, VT, LHS);
9677 RHS = DAG.getNode(ISD::BITCAST, dl, VT, RHS);
9680 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, LHS, RHS,
9681 DAG.getConstant(MaskValue, MVT::i32));
9682 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
9685 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
9686 SDValue BlendOp = LowerVSELECTtoBlend(Op, Subtarget, DAG);
9687 if (BlendOp.getNode())
9690 // Some types for vselect were previously set to Expand, not Legal or
9691 // Custom. Return an empty SDValue so we fall-through to Expand, after
9692 // the Custom lowering phase.
9693 MVT VT = Op.getSimpleValueType();
9694 switch (VT.SimpleTy) {
9702 // We couldn't create a "Blend with immediate" node.
9703 // This node should still be legal, but we'll have to emit a blendv*
9708 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
9709 MVT VT = Op.getSimpleValueType();
9712 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
9715 if (VT.getSizeInBits() == 8) {
9716 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
9717 Op.getOperand(0), Op.getOperand(1));
9718 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
9719 DAG.getValueType(VT));
9720 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
9723 if (VT.getSizeInBits() == 16) {
9724 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9725 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
9727 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
9728 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
9729 DAG.getNode(ISD::BITCAST, dl,
9733 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
9734 Op.getOperand(0), Op.getOperand(1));
9735 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
9736 DAG.getValueType(VT));
9737 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
9740 if (VT == MVT::f32) {
9741 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
9742 // the result back to FR32 register. It's only worth matching if the
9743 // result has a single use which is a store or a bitcast to i32. And in
9744 // the case of a store, it's not worth it if the index is a constant 0,
9745 // because a MOVSSmr can be used instead, which is smaller and faster.
9746 if (!Op.hasOneUse())
9748 SDNode *User = *Op.getNode()->use_begin();
9749 if ((User->getOpcode() != ISD::STORE ||
9750 (isa<ConstantSDNode>(Op.getOperand(1)) &&
9751 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
9752 (User->getOpcode() != ISD::BITCAST ||
9753 User->getValueType(0) != MVT::i32))
9755 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
9756 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
9759 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
9762 if (VT == MVT::i32 || VT == MVT::i64) {
9763 // ExtractPS/pextrq works with constant index.
9764 if (isa<ConstantSDNode>(Op.getOperand(1)))
9770 /// Extract one bit from mask vector, like v16i1 or v8i1.
9771 /// AVX-512 feature.
9773 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
9774 SDValue Vec = Op.getOperand(0);
9776 MVT VecVT = Vec.getSimpleValueType();
9777 SDValue Idx = Op.getOperand(1);
9778 MVT EltVT = Op.getSimpleValueType();
9780 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
9782 // variable index can't be handled in mask registers,
9783 // extend vector to VR512
9784 if (!isa<ConstantSDNode>(Idx)) {
9785 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
9786 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
9787 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
9788 ExtVT.getVectorElementType(), Ext, Idx);
9789 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
9792 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
9793 const TargetRegisterClass* rc = getRegClassFor(VecVT);
9794 unsigned MaxSift = rc->getSize()*8 - 1;
9795 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
9796 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
9797 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
9798 DAG.getConstant(MaxSift, MVT::i8));
9799 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
9800 DAG.getIntPtrConstant(0));
9804 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
9805 SelectionDAG &DAG) const {
9807 SDValue Vec = Op.getOperand(0);
9808 MVT VecVT = Vec.getSimpleValueType();
9809 SDValue Idx = Op.getOperand(1);
9811 if (Op.getSimpleValueType() == MVT::i1)
9812 return ExtractBitFromMaskVector(Op, DAG);
9814 if (!isa<ConstantSDNode>(Idx)) {
9815 if (VecVT.is512BitVector() ||
9816 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
9817 VecVT.getVectorElementType().getSizeInBits() == 32)) {
9820 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
9821 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
9822 MaskEltVT.getSizeInBits());
9824 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
9825 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
9826 getZeroVector(MaskVT, Subtarget, DAG, dl),
9827 Idx, DAG.getConstant(0, getPointerTy()));
9828 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
9829 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
9830 Perm, DAG.getConstant(0, getPointerTy()));
9835 // If this is a 256-bit vector result, first extract the 128-bit vector and
9836 // then extract the element from the 128-bit vector.
9837 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
9839 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
9840 // Get the 128-bit vector.
9841 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
9842 MVT EltVT = VecVT.getVectorElementType();
9844 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
9846 //if (IdxVal >= NumElems/2)
9847 // IdxVal -= NumElems/2;
9848 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
9849 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
9850 DAG.getConstant(IdxVal, MVT::i32));
9853 assert(VecVT.is128BitVector() && "Unexpected vector length");
9855 if (Subtarget->hasSSE41()) {
9856 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
9861 MVT VT = Op.getSimpleValueType();
9862 // TODO: handle v16i8.
9863 if (VT.getSizeInBits() == 16) {
9864 SDValue Vec = Op.getOperand(0);
9865 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9867 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
9868 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
9869 DAG.getNode(ISD::BITCAST, dl,
9872 // Transform it so it match pextrw which produces a 32-bit result.
9873 MVT EltVT = MVT::i32;
9874 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
9875 Op.getOperand(0), Op.getOperand(1));
9876 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
9877 DAG.getValueType(VT));
9878 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
9881 if (VT.getSizeInBits() == 32) {
9882 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9886 // SHUFPS the element to the lowest double word, then movss.
9887 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
9888 MVT VVT = Op.getOperand(0).getSimpleValueType();
9889 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
9890 DAG.getUNDEF(VVT), Mask);
9891 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
9892 DAG.getIntPtrConstant(0));
9895 if (VT.getSizeInBits() == 64) {
9896 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
9897 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
9898 // to match extract_elt for f64.
9899 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9903 // UNPCKHPD the element to the lowest double word, then movsd.
9904 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
9905 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
9906 int Mask[2] = { 1, -1 };
9907 MVT VVT = Op.getOperand(0).getSimpleValueType();
9908 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
9909 DAG.getUNDEF(VVT), Mask);
9910 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
9911 DAG.getIntPtrConstant(0));
9917 static SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
9918 MVT VT = Op.getSimpleValueType();
9919 MVT EltVT = VT.getVectorElementType();
9922 SDValue N0 = Op.getOperand(0);
9923 SDValue N1 = Op.getOperand(1);
9924 SDValue N2 = Op.getOperand(2);
9926 if (!VT.is128BitVector())
9929 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
9930 isa<ConstantSDNode>(N2)) {
9932 if (VT == MVT::v8i16)
9933 Opc = X86ISD::PINSRW;
9934 else if (VT == MVT::v16i8)
9935 Opc = X86ISD::PINSRB;
9937 Opc = X86ISD::PINSRB;
9939 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
9941 if (N1.getValueType() != MVT::i32)
9942 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
9943 if (N2.getValueType() != MVT::i32)
9944 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
9945 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
9948 if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
9949 // Bits [7:6] of the constant are the source select. This will always be
9950 // zero here. The DAG Combiner may combine an extract_elt index into these
9951 // bits. For example (insert (extract, 3), 2) could be matched by putting
9952 // the '3' into bits [7:6] of X86ISD::INSERTPS.
9953 // Bits [5:4] of the constant are the destination select. This is the
9954 // value of the incoming immediate.
9955 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
9956 // combine either bitwise AND or insert of float 0.0 to set these bits.
9957 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
9958 // Create this as a scalar to vector..
9959 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
9960 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
9963 if ((EltVT == MVT::i32 || EltVT == MVT::i64) && isa<ConstantSDNode>(N2)) {
9964 // PINSR* works with constant index.
9970 /// Insert one bit to mask vector, like v16i1 or v8i1.
9971 /// AVX-512 feature.
9973 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
9975 SDValue Vec = Op.getOperand(0);
9976 SDValue Elt = Op.getOperand(1);
9977 SDValue Idx = Op.getOperand(2);
9978 MVT VecVT = Vec.getSimpleValueType();
9980 if (!isa<ConstantSDNode>(Idx)) {
9981 // Non constant index. Extend source and destination,
9982 // insert element and then truncate the result.
9983 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
9984 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
9985 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
9986 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
9987 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
9988 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
9991 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
9992 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
9993 if (Vec.getOpcode() == ISD::UNDEF)
9994 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
9995 DAG.getConstant(IdxVal, MVT::i8));
9996 const TargetRegisterClass* rc = getRegClassFor(VecVT);
9997 unsigned MaxSift = rc->getSize()*8 - 1;
9998 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
9999 DAG.getConstant(MaxSift, MVT::i8));
10000 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
10001 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
10002 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
10005 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
10006 MVT VT = Op.getSimpleValueType();
10007 MVT EltVT = VT.getVectorElementType();
10009 if (EltVT == MVT::i1)
10010 return InsertBitToMaskVector(Op, DAG);
10013 SDValue N0 = Op.getOperand(0);
10014 SDValue N1 = Op.getOperand(1);
10015 SDValue N2 = Op.getOperand(2);
10017 // If this is a 256-bit vector result, first extract the 128-bit vector,
10018 // insert the element into the extracted half and then place it back.
10019 if (VT.is256BitVector() || VT.is512BitVector()) {
10020 if (!isa<ConstantSDNode>(N2))
10023 // Get the desired 128-bit vector half.
10024 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
10025 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
10027 // Insert the element into the desired half.
10028 unsigned NumEltsIn128 = 128/EltVT.getSizeInBits();
10029 unsigned IdxIn128 = IdxVal - (IdxVal/NumEltsIn128) * NumEltsIn128;
10031 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
10032 DAG.getConstant(IdxIn128, MVT::i32));
10034 // Insert the changed part back to the 256-bit vector
10035 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
10038 if (Subtarget->hasSSE41())
10039 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
10041 if (EltVT == MVT::i8)
10044 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
10045 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
10046 // as its second argument.
10047 if (N1.getValueType() != MVT::i32)
10048 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10049 if (N2.getValueType() != MVT::i32)
10050 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
10051 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
10056 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
10058 MVT OpVT = Op.getSimpleValueType();
10060 // If this is a 256-bit vector result, first insert into a 128-bit
10061 // vector and then insert into the 256-bit vector.
10062 if (!OpVT.is128BitVector()) {
10063 // Insert into a 128-bit vector.
10064 unsigned SizeFactor = OpVT.getSizeInBits()/128;
10065 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
10066 OpVT.getVectorNumElements() / SizeFactor);
10068 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
10070 // Insert the 128-bit vector.
10071 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
10074 if (OpVT == MVT::v1i64 &&
10075 Op.getOperand(0).getValueType() == MVT::i64)
10076 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
10078 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
10079 assert(OpVT.is128BitVector() && "Expected an SSE type!");
10080 return DAG.getNode(ISD::BITCAST, dl, OpVT,
10081 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
10084 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
10085 // a simple subregister reference or explicit instructions to grab
10086 // upper bits of a vector.
10087 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10088 SelectionDAG &DAG) {
10090 SDValue In = Op.getOperand(0);
10091 SDValue Idx = Op.getOperand(1);
10092 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10093 MVT ResVT = Op.getSimpleValueType();
10094 MVT InVT = In.getSimpleValueType();
10096 if (Subtarget->hasFp256()) {
10097 if (ResVT.is128BitVector() &&
10098 (InVT.is256BitVector() || InVT.is512BitVector()) &&
10099 isa<ConstantSDNode>(Idx)) {
10100 return Extract128BitVector(In, IdxVal, DAG, dl);
10102 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
10103 isa<ConstantSDNode>(Idx)) {
10104 return Extract256BitVector(In, IdxVal, DAG, dl);
10110 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
10111 // simple superregister reference or explicit instructions to insert
10112 // the upper bits of a vector.
10113 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10114 SelectionDAG &DAG) {
10115 if (Subtarget->hasFp256()) {
10116 SDLoc dl(Op.getNode());
10117 SDValue Vec = Op.getNode()->getOperand(0);
10118 SDValue SubVec = Op.getNode()->getOperand(1);
10119 SDValue Idx = Op.getNode()->getOperand(2);
10121 if ((Op.getNode()->getSimpleValueType(0).is256BitVector() ||
10122 Op.getNode()->getSimpleValueType(0).is512BitVector()) &&
10123 SubVec.getNode()->getSimpleValueType(0).is128BitVector() &&
10124 isa<ConstantSDNode>(Idx)) {
10125 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10126 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
10129 if (Op.getNode()->getSimpleValueType(0).is512BitVector() &&
10130 SubVec.getNode()->getSimpleValueType(0).is256BitVector() &&
10131 isa<ConstantSDNode>(Idx)) {
10132 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10133 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
10139 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
10140 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
10141 // one of the above mentioned nodes. It has to be wrapped because otherwise
10142 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
10143 // be used to form addressing mode. These wrapped nodes will be selected
10146 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
10147 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
10149 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10150 // global base reg.
10151 unsigned char OpFlag = 0;
10152 unsigned WrapperKind = X86ISD::Wrapper;
10153 CodeModel::Model M = DAG.getTarget().getCodeModel();
10155 if (Subtarget->isPICStyleRIPRel() &&
10156 (M == CodeModel::Small || M == CodeModel::Kernel))
10157 WrapperKind = X86ISD::WrapperRIP;
10158 else if (Subtarget->isPICStyleGOT())
10159 OpFlag = X86II::MO_GOTOFF;
10160 else if (Subtarget->isPICStyleStubPIC())
10161 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10163 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
10164 CP->getAlignment(),
10165 CP->getOffset(), OpFlag);
10167 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10168 // With PIC, the address is actually $g + Offset.
10170 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10171 DAG.getNode(X86ISD::GlobalBaseReg,
10172 SDLoc(), getPointerTy()),
10179 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
10180 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
10182 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10183 // global base reg.
10184 unsigned char OpFlag = 0;
10185 unsigned WrapperKind = X86ISD::Wrapper;
10186 CodeModel::Model M = DAG.getTarget().getCodeModel();
10188 if (Subtarget->isPICStyleRIPRel() &&
10189 (M == CodeModel::Small || M == CodeModel::Kernel))
10190 WrapperKind = X86ISD::WrapperRIP;
10191 else if (Subtarget->isPICStyleGOT())
10192 OpFlag = X86II::MO_GOTOFF;
10193 else if (Subtarget->isPICStyleStubPIC())
10194 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10196 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
10199 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10201 // With PIC, the address is actually $g + Offset.
10203 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10204 DAG.getNode(X86ISD::GlobalBaseReg,
10205 SDLoc(), getPointerTy()),
10212 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
10213 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
10215 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10216 // global base reg.
10217 unsigned char OpFlag = 0;
10218 unsigned WrapperKind = X86ISD::Wrapper;
10219 CodeModel::Model M = DAG.getTarget().getCodeModel();
10221 if (Subtarget->isPICStyleRIPRel() &&
10222 (M == CodeModel::Small || M == CodeModel::Kernel)) {
10223 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
10224 OpFlag = X86II::MO_GOTPCREL;
10225 WrapperKind = X86ISD::WrapperRIP;
10226 } else if (Subtarget->isPICStyleGOT()) {
10227 OpFlag = X86II::MO_GOT;
10228 } else if (Subtarget->isPICStyleStubPIC()) {
10229 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
10230 } else if (Subtarget->isPICStyleStubNoDynamic()) {
10231 OpFlag = X86II::MO_DARWIN_NONLAZY;
10234 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
10237 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10239 // With PIC, the address is actually $g + Offset.
10240 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
10241 !Subtarget->is64Bit()) {
10242 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10243 DAG.getNode(X86ISD::GlobalBaseReg,
10244 SDLoc(), getPointerTy()),
10248 // For symbols that require a load from a stub to get the address, emit the
10250 if (isGlobalStubReference(OpFlag))
10251 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
10252 MachinePointerInfo::getGOT(), false, false, false, 0);
10258 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
10259 // Create the TargetBlockAddressAddress node.
10260 unsigned char OpFlags =
10261 Subtarget->ClassifyBlockAddressReference();
10262 CodeModel::Model M = DAG.getTarget().getCodeModel();
10263 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
10264 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
10266 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
10269 if (Subtarget->isPICStyleRIPRel() &&
10270 (M == CodeModel::Small || M == CodeModel::Kernel))
10271 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
10273 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
10275 // With PIC, the address is actually $g + Offset.
10276 if (isGlobalRelativeToPICBase(OpFlags)) {
10277 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
10278 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
10286 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
10287 int64_t Offset, SelectionDAG &DAG) const {
10288 // Create the TargetGlobalAddress node, folding in the constant
10289 // offset if it is legal.
10290 unsigned char OpFlags =
10291 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
10292 CodeModel::Model M = DAG.getTarget().getCodeModel();
10294 if (OpFlags == X86II::MO_NO_FLAG &&
10295 X86::isOffsetSuitableForCodeModel(Offset, M)) {
10296 // A direct static reference to a global.
10297 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
10300 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
10303 if (Subtarget->isPICStyleRIPRel() &&
10304 (M == CodeModel::Small || M == CodeModel::Kernel))
10305 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
10307 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
10309 // With PIC, the address is actually $g + Offset.
10310 if (isGlobalRelativeToPICBase(OpFlags)) {
10311 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
10312 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
10316 // For globals that require a load from a stub to get the address, emit the
10318 if (isGlobalStubReference(OpFlags))
10319 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
10320 MachinePointerInfo::getGOT(), false, false, false, 0);
10322 // If there was a non-zero offset that we didn't fold, create an explicit
10323 // addition for it.
10325 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
10326 DAG.getConstant(Offset, getPointerTy()));
10332 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
10333 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
10334 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
10335 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
10339 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
10340 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
10341 unsigned char OperandFlags, bool LocalDynamic = false) {
10342 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10343 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10345 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
10346 GA->getValueType(0),
10350 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
10354 SDValue Ops[] = { Chain, TGA, *InFlag };
10355 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
10357 SDValue Ops[] = { Chain, TGA };
10358 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
10361 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
10362 MFI->setAdjustsStack(true);
10364 SDValue Flag = Chain.getValue(1);
10365 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
10368 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
10370 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10373 SDLoc dl(GA); // ? function entry point might be better
10374 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
10375 DAG.getNode(X86ISD::GlobalBaseReg,
10376 SDLoc(), PtrVT), InFlag);
10377 InFlag = Chain.getValue(1);
10379 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
10382 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
10384 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10386 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
10387 X86::RAX, X86II::MO_TLSGD);
10390 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
10396 // Get the start address of the TLS block for this module.
10397 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
10398 .getInfo<X86MachineFunctionInfo>();
10399 MFI->incNumLocalDynamicTLSAccesses();
10403 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
10404 X86II::MO_TLSLD, /*LocalDynamic=*/true);
10407 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
10408 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
10409 InFlag = Chain.getValue(1);
10410 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
10411 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
10414 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
10418 unsigned char OperandFlags = X86II::MO_DTPOFF;
10419 unsigned WrapperKind = X86ISD::Wrapper;
10420 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
10421 GA->getValueType(0),
10422 GA->getOffset(), OperandFlags);
10423 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
10425 // Add x@dtpoff with the base.
10426 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
10429 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
10430 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10431 const EVT PtrVT, TLSModel::Model model,
10432 bool is64Bit, bool isPIC) {
10435 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
10436 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
10437 is64Bit ? 257 : 256));
10439 SDValue ThreadPointer =
10440 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0),
10441 MachinePointerInfo(Ptr), false, false, false, 0);
10443 unsigned char OperandFlags = 0;
10444 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
10446 unsigned WrapperKind = X86ISD::Wrapper;
10447 if (model == TLSModel::LocalExec) {
10448 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
10449 } else if (model == TLSModel::InitialExec) {
10451 OperandFlags = X86II::MO_GOTTPOFF;
10452 WrapperKind = X86ISD::WrapperRIP;
10454 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
10457 llvm_unreachable("Unexpected model");
10460 // emit "addl x@ntpoff,%eax" (local exec)
10461 // or "addl x@indntpoff,%eax" (initial exec)
10462 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
10464 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
10465 GA->getOffset(), OperandFlags);
10466 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
10468 if (model == TLSModel::InitialExec) {
10469 if (isPIC && !is64Bit) {
10470 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
10471 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
10475 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
10476 MachinePointerInfo::getGOT(), false, false, false, 0);
10479 // The address of the thread local variable is the add of the thread
10480 // pointer with the offset of the variable.
10481 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
10485 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
10487 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
10488 const GlobalValue *GV = GA->getGlobal();
10490 if (Subtarget->isTargetELF()) {
10491 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
10494 case TLSModel::GeneralDynamic:
10495 if (Subtarget->is64Bit())
10496 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
10497 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
10498 case TLSModel::LocalDynamic:
10499 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
10500 Subtarget->is64Bit());
10501 case TLSModel::InitialExec:
10502 case TLSModel::LocalExec:
10503 return LowerToTLSExecModel(
10504 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
10505 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
10507 llvm_unreachable("Unknown TLS model.");
10510 if (Subtarget->isTargetDarwin()) {
10511 // Darwin only has one model of TLS. Lower to that.
10512 unsigned char OpFlag = 0;
10513 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
10514 X86ISD::WrapperRIP : X86ISD::Wrapper;
10516 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10517 // global base reg.
10518 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
10519 !Subtarget->is64Bit();
10521 OpFlag = X86II::MO_TLVP_PIC_BASE;
10523 OpFlag = X86II::MO_TLVP;
10525 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
10526 GA->getValueType(0),
10527 GA->getOffset(), OpFlag);
10528 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10530 // With PIC32, the address is actually $g + Offset.
10532 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10533 DAG.getNode(X86ISD::GlobalBaseReg,
10534 SDLoc(), getPointerTy()),
10537 // Lowering the machine isd will make sure everything is in the right
10539 SDValue Chain = DAG.getEntryNode();
10540 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10541 SDValue Args[] = { Chain, Offset };
10542 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
10544 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
10545 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10546 MFI->setAdjustsStack(true);
10548 // And our return value (tls address) is in the standard call return value
10550 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10551 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
10552 Chain.getValue(1));
10555 if (Subtarget->isTargetKnownWindowsMSVC() ||
10556 Subtarget->isTargetWindowsGNU()) {
10557 // Just use the implicit TLS architecture
10558 // Need to generate someting similar to:
10559 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
10561 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
10562 // mov rcx, qword [rdx+rcx*8]
10563 // mov eax, .tls$:tlsvar
10564 // [rax+rcx] contains the address
10565 // Windows 64bit: gs:0x58
10566 // Windows 32bit: fs:__tls_array
10569 SDValue Chain = DAG.getEntryNode();
10571 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
10572 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
10573 // use its literal value of 0x2C.
10574 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
10575 ? Type::getInt8PtrTy(*DAG.getContext(),
10577 : Type::getInt32PtrTy(*DAG.getContext(),
10581 Subtarget->is64Bit()
10582 ? DAG.getIntPtrConstant(0x58)
10583 : (Subtarget->isTargetWindowsGNU()
10584 ? DAG.getIntPtrConstant(0x2C)
10585 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
10587 SDValue ThreadPointer =
10588 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
10589 MachinePointerInfo(Ptr), false, false, false, 0);
10591 // Load the _tls_index variable
10592 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
10593 if (Subtarget->is64Bit())
10594 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
10595 IDX, MachinePointerInfo(), MVT::i32,
10598 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
10599 false, false, false, 0);
10601 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
10603 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
10605 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
10606 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
10607 false, false, false, 0);
10609 // Get the offset of start of .tls section
10610 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
10611 GA->getValueType(0),
10612 GA->getOffset(), X86II::MO_SECREL);
10613 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
10615 // The address of the thread local variable is the add of the thread
10616 // pointer with the offset of the variable.
10617 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
10620 llvm_unreachable("TLS not implemented for this target.");
10623 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
10624 /// and take a 2 x i32 value to shift plus a shift amount.
10625 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
10626 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
10627 MVT VT = Op.getSimpleValueType();
10628 unsigned VTBits = VT.getSizeInBits();
10630 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
10631 SDValue ShOpLo = Op.getOperand(0);
10632 SDValue ShOpHi = Op.getOperand(1);
10633 SDValue ShAmt = Op.getOperand(2);
10634 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
10635 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
10637 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
10638 DAG.getConstant(VTBits - 1, MVT::i8));
10639 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
10640 DAG.getConstant(VTBits - 1, MVT::i8))
10641 : DAG.getConstant(0, VT);
10643 SDValue Tmp2, Tmp3;
10644 if (Op.getOpcode() == ISD::SHL_PARTS) {
10645 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
10646 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
10648 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
10649 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
10652 // If the shift amount is larger or equal than the width of a part we can't
10653 // rely on the results of shld/shrd. Insert a test and select the appropriate
10654 // values for large shift amounts.
10655 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
10656 DAG.getConstant(VTBits, MVT::i8));
10657 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
10658 AndNode, DAG.getConstant(0, MVT::i8));
10661 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10662 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
10663 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
10665 if (Op.getOpcode() == ISD::SHL_PARTS) {
10666 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
10667 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
10669 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
10670 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
10673 SDValue Ops[2] = { Lo, Hi };
10674 return DAG.getMergeValues(Ops, dl);
10677 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
10678 SelectionDAG &DAG) const {
10679 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
10681 if (SrcVT.isVector())
10684 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
10685 "Unknown SINT_TO_FP to lower!");
10687 // These are really Legal; return the operand so the caller accepts it as
10689 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
10691 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
10692 Subtarget->is64Bit()) {
10697 unsigned Size = SrcVT.getSizeInBits()/8;
10698 MachineFunction &MF = DAG.getMachineFunction();
10699 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
10700 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
10701 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
10703 MachinePointerInfo::getFixedStack(SSFI),
10705 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
10708 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
10710 SelectionDAG &DAG) const {
10714 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
10716 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
10718 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
10720 unsigned ByteSize = SrcVT.getSizeInBits()/8;
10722 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
10723 MachineMemOperand *MMO;
10725 int SSFI = FI->getIndex();
10727 DAG.getMachineFunction()
10728 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
10729 MachineMemOperand::MOLoad, ByteSize, ByteSize);
10731 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
10732 StackSlot = StackSlot.getOperand(1);
10734 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
10735 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
10737 Tys, Ops, SrcVT, MMO);
10740 Chain = Result.getValue(1);
10741 SDValue InFlag = Result.getValue(2);
10743 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
10744 // shouldn't be necessary except that RFP cannot be live across
10745 // multiple blocks. When stackifier is fixed, they can be uncoupled.
10746 MachineFunction &MF = DAG.getMachineFunction();
10747 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
10748 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
10749 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
10750 Tys = DAG.getVTList(MVT::Other);
10752 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
10754 MachineMemOperand *MMO =
10755 DAG.getMachineFunction()
10756 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
10757 MachineMemOperand::MOStore, SSFISize, SSFISize);
10759 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
10760 Ops, Op.getValueType(), MMO);
10761 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
10762 MachinePointerInfo::getFixedStack(SSFI),
10763 false, false, false, 0);
10769 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
10770 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
10771 SelectionDAG &DAG) const {
10772 // This algorithm is not obvious. Here it is what we're trying to output:
10775 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
10776 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
10778 haddpd %xmm0, %xmm0
10780 pshufd $0x4e, %xmm0, %xmm1
10786 LLVMContext *Context = DAG.getContext();
10788 // Build some magic constants.
10789 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
10790 Constant *C0 = ConstantDataVector::get(*Context, CV0);
10791 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
10793 SmallVector<Constant*,2> CV1;
10795 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
10796 APInt(64, 0x4330000000000000ULL))));
10798 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
10799 APInt(64, 0x4530000000000000ULL))));
10800 Constant *C1 = ConstantVector::get(CV1);
10801 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
10803 // Load the 64-bit value into an XMM register.
10804 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
10806 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
10807 MachinePointerInfo::getConstantPool(),
10808 false, false, false, 16);
10809 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
10810 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
10813 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
10814 MachinePointerInfo::getConstantPool(),
10815 false, false, false, 16);
10816 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
10817 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
10820 if (Subtarget->hasSSE3()) {
10821 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
10822 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
10824 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
10825 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
10827 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
10828 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
10832 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
10833 DAG.getIntPtrConstant(0));
10836 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
10837 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
10838 SelectionDAG &DAG) const {
10840 // FP constant to bias correct the final result.
10841 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
10844 // Load the 32-bit value into an XMM register.
10845 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
10848 // Zero out the upper parts of the register.
10849 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
10851 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
10852 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
10853 DAG.getIntPtrConstant(0));
10855 // Or the load with the bias.
10856 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
10857 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
10858 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
10859 MVT::v2f64, Load)),
10860 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
10861 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
10862 MVT::v2f64, Bias)));
10863 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
10864 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
10865 DAG.getIntPtrConstant(0));
10867 // Subtract the bias.
10868 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
10870 // Handle final rounding.
10871 EVT DestVT = Op.getValueType();
10873 if (DestVT.bitsLT(MVT::f64))
10874 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
10875 DAG.getIntPtrConstant(0));
10876 if (DestVT.bitsGT(MVT::f64))
10877 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
10879 // Handle final rounding.
10883 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
10884 SelectionDAG &DAG) const {
10885 SDValue N0 = Op.getOperand(0);
10886 MVT SVT = N0.getSimpleValueType();
10889 assert((SVT == MVT::v4i8 || SVT == MVT::v4i16 ||
10890 SVT == MVT::v8i8 || SVT == MVT::v8i16) &&
10891 "Custom UINT_TO_FP is not supported!");
10893 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
10894 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
10895 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
10898 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
10899 SelectionDAG &DAG) const {
10900 SDValue N0 = Op.getOperand(0);
10903 if (Op.getValueType().isVector())
10904 return lowerUINT_TO_FP_vec(Op, DAG);
10906 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
10907 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
10908 // the optimization here.
10909 if (DAG.SignBitIsZero(N0))
10910 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
10912 MVT SrcVT = N0.getSimpleValueType();
10913 MVT DstVT = Op.getSimpleValueType();
10914 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
10915 return LowerUINT_TO_FP_i64(Op, DAG);
10916 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
10917 return LowerUINT_TO_FP_i32(Op, DAG);
10918 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
10921 // Make a 64-bit buffer, and use it to build an FILD.
10922 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
10923 if (SrcVT == MVT::i32) {
10924 SDValue WordOff = DAG.getConstant(4, getPointerTy());
10925 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
10926 getPointerTy(), StackSlot, WordOff);
10927 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
10928 StackSlot, MachinePointerInfo(),
10930 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
10931 OffsetSlot, MachinePointerInfo(),
10933 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
10937 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
10938 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
10939 StackSlot, MachinePointerInfo(),
10941 // For i64 source, we need to add the appropriate power of 2 if the input
10942 // was negative. This is the same as the optimization in
10943 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
10944 // we must be careful to do the computation in x87 extended precision, not
10945 // in SSE. (The generic code can't know it's OK to do this, or how to.)
10946 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
10947 MachineMemOperand *MMO =
10948 DAG.getMachineFunction()
10949 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
10950 MachineMemOperand::MOLoad, 8, 8);
10952 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
10953 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
10954 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
10957 APInt FF(32, 0x5F800000ULL);
10959 // Check whether the sign bit is set.
10960 SDValue SignSet = DAG.getSetCC(dl,
10961 getSetCCResultType(*DAG.getContext(), MVT::i64),
10962 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
10965 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
10966 SDValue FudgePtr = DAG.getConstantPool(
10967 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
10970 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
10971 SDValue Zero = DAG.getIntPtrConstant(0);
10972 SDValue Four = DAG.getIntPtrConstant(4);
10973 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
10975 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
10977 // Load the value out, extending it from f32 to f80.
10978 // FIXME: Avoid the extend by constructing the right constant pool?
10979 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
10980 FudgePtr, MachinePointerInfo::getConstantPool(),
10981 MVT::f32, false, false, 4);
10982 // Extend everything to 80 bits to force it to be done on x87.
10983 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
10984 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
10987 std::pair<SDValue,SDValue>
10988 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
10989 bool IsSigned, bool IsReplace) const {
10992 EVT DstTy = Op.getValueType();
10994 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
10995 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
10999 assert(DstTy.getSimpleVT() <= MVT::i64 &&
11000 DstTy.getSimpleVT() >= MVT::i16 &&
11001 "Unknown FP_TO_INT to lower!");
11003 // These are really Legal.
11004 if (DstTy == MVT::i32 &&
11005 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11006 return std::make_pair(SDValue(), SDValue());
11007 if (Subtarget->is64Bit() &&
11008 DstTy == MVT::i64 &&
11009 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11010 return std::make_pair(SDValue(), SDValue());
11012 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
11013 // stack slot, or into the FTOL runtime function.
11014 MachineFunction &MF = DAG.getMachineFunction();
11015 unsigned MemSize = DstTy.getSizeInBits()/8;
11016 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11017 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11020 if (!IsSigned && isIntegerTypeFTOL(DstTy))
11021 Opc = X86ISD::WIN_FTOL;
11023 switch (DstTy.getSimpleVT().SimpleTy) {
11024 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
11025 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
11026 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
11027 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
11030 SDValue Chain = DAG.getEntryNode();
11031 SDValue Value = Op.getOperand(0);
11032 EVT TheVT = Op.getOperand(0).getValueType();
11033 // FIXME This causes a redundant load/store if the SSE-class value is already
11034 // in memory, such as if it is on the callstack.
11035 if (isScalarFPTypeInSSEReg(TheVT)) {
11036 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
11037 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
11038 MachinePointerInfo::getFixedStack(SSFI),
11040 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
11042 Chain, StackSlot, DAG.getValueType(TheVT)
11045 MachineMemOperand *MMO =
11046 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11047 MachineMemOperand::MOLoad, MemSize, MemSize);
11048 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
11049 Chain = Value.getValue(1);
11050 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11051 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11054 MachineMemOperand *MMO =
11055 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11056 MachineMemOperand::MOStore, MemSize, MemSize);
11058 if (Opc != X86ISD::WIN_FTOL) {
11059 // Build the FP_TO_INT*_IN_MEM
11060 SDValue Ops[] = { Chain, Value, StackSlot };
11061 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
11063 return std::make_pair(FIST, StackSlot);
11065 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
11066 DAG.getVTList(MVT::Other, MVT::Glue),
11068 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
11069 MVT::i32, ftol.getValue(1));
11070 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
11071 MVT::i32, eax.getValue(2));
11072 SDValue Ops[] = { eax, edx };
11073 SDValue pair = IsReplace
11074 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
11075 : DAG.getMergeValues(Ops, DL);
11076 return std::make_pair(pair, SDValue());
11080 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
11081 const X86Subtarget *Subtarget) {
11082 MVT VT = Op->getSimpleValueType(0);
11083 SDValue In = Op->getOperand(0);
11084 MVT InVT = In.getSimpleValueType();
11087 // Optimize vectors in AVX mode:
11090 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
11091 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
11092 // Concat upper and lower parts.
11095 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
11096 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
11097 // Concat upper and lower parts.
11100 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
11101 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
11102 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
11105 if (Subtarget->hasInt256())
11106 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
11108 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
11109 SDValue Undef = DAG.getUNDEF(InVT);
11110 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
11111 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11112 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11114 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
11115 VT.getVectorNumElements()/2);
11117 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
11118 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
11120 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
11123 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
11124 SelectionDAG &DAG) {
11125 MVT VT = Op->getSimpleValueType(0);
11126 SDValue In = Op->getOperand(0);
11127 MVT InVT = In.getSimpleValueType();
11129 unsigned int NumElts = VT.getVectorNumElements();
11130 if (NumElts != 8 && NumElts != 16)
11133 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
11134 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
11136 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
11137 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11138 // Now we have only mask extension
11139 assert(InVT.getVectorElementType() == MVT::i1);
11140 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
11141 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
11142 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
11143 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
11144 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
11145 MachinePointerInfo::getConstantPool(),
11146 false, false, false, Alignment);
11148 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
11149 if (VT.is512BitVector())
11151 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
11154 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
11155 SelectionDAG &DAG) {
11156 if (Subtarget->hasFp256()) {
11157 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
11165 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
11166 SelectionDAG &DAG) {
11168 MVT VT = Op.getSimpleValueType();
11169 SDValue In = Op.getOperand(0);
11170 MVT SVT = In.getSimpleValueType();
11172 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
11173 return LowerZERO_EXTEND_AVX512(Op, DAG);
11175 if (Subtarget->hasFp256()) {
11176 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
11181 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
11182 VT.getVectorNumElements() != SVT.getVectorNumElements());
11186 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
11188 MVT VT = Op.getSimpleValueType();
11189 SDValue In = Op.getOperand(0);
11190 MVT InVT = In.getSimpleValueType();
11192 if (VT == MVT::i1) {
11193 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
11194 "Invalid scalar TRUNCATE operation");
11195 if (InVT == MVT::i32)
11197 if (InVT.getSizeInBits() == 64)
11198 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::i32, In);
11199 else if (InVT.getSizeInBits() < 32)
11200 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
11201 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
11203 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
11204 "Invalid TRUNCATE operation");
11206 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
11207 if (VT.getVectorElementType().getSizeInBits() >=8)
11208 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
11210 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
11211 unsigned NumElts = InVT.getVectorNumElements();
11212 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
11213 if (InVT.getSizeInBits() < 512) {
11214 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
11215 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
11219 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
11220 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
11221 SDValue CP = DAG.getConstantPool(C, getPointerTy());
11222 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
11223 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
11224 MachinePointerInfo::getConstantPool(),
11225 false, false, false, Alignment);
11226 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
11227 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
11228 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
11231 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
11232 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
11233 if (Subtarget->hasInt256()) {
11234 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
11235 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
11236 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
11238 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
11239 DAG.getIntPtrConstant(0));
11242 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11243 DAG.getIntPtrConstant(0));
11244 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11245 DAG.getIntPtrConstant(2));
11246 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
11247 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
11248 static const int ShufMask[] = {0, 2, 4, 6};
11249 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
11252 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
11253 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
11254 if (Subtarget->hasInt256()) {
11255 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
11257 SmallVector<SDValue,32> pshufbMask;
11258 for (unsigned i = 0; i < 2; ++i) {
11259 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
11260 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
11261 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
11262 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
11263 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
11264 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
11265 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
11266 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
11267 for (unsigned j = 0; j < 8; ++j)
11268 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
11270 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
11271 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
11272 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
11274 static const int ShufMask[] = {0, 2, -1, -1};
11275 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
11277 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11278 DAG.getIntPtrConstant(0));
11279 return DAG.getNode(ISD::BITCAST, DL, VT, In);
11282 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
11283 DAG.getIntPtrConstant(0));
11285 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
11286 DAG.getIntPtrConstant(4));
11288 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
11289 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
11291 // The PSHUFB mask:
11292 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
11293 -1, -1, -1, -1, -1, -1, -1, -1};
11295 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
11296 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
11297 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
11299 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
11300 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
11302 // The MOVLHPS Mask:
11303 static const int ShufMask2[] = {0, 1, 4, 5};
11304 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
11305 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
11308 // Handle truncation of V256 to V128 using shuffles.
11309 if (!VT.is128BitVector() || !InVT.is256BitVector())
11312 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
11314 unsigned NumElems = VT.getVectorNumElements();
11315 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
11317 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
11318 // Prepare truncation shuffle mask
11319 for (unsigned i = 0; i != NumElems; ++i)
11320 MaskVec[i] = i * 2;
11321 SDValue V = DAG.getVectorShuffle(NVT, DL,
11322 DAG.getNode(ISD::BITCAST, DL, NVT, In),
11323 DAG.getUNDEF(NVT), &MaskVec[0]);
11324 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
11325 DAG.getIntPtrConstant(0));
11328 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
11329 SelectionDAG &DAG) const {
11330 assert(!Op.getSimpleValueType().isVector());
11332 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
11333 /*IsSigned=*/ true, /*IsReplace=*/ false);
11334 SDValue FIST = Vals.first, StackSlot = Vals.second;
11335 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
11336 if (!FIST.getNode()) return Op;
11338 if (StackSlot.getNode())
11339 // Load the result.
11340 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
11341 FIST, StackSlot, MachinePointerInfo(),
11342 false, false, false, 0);
11344 // The node is the result.
11348 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
11349 SelectionDAG &DAG) const {
11350 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
11351 /*IsSigned=*/ false, /*IsReplace=*/ false);
11352 SDValue FIST = Vals.first, StackSlot = Vals.second;
11353 assert(FIST.getNode() && "Unexpected failure");
11355 if (StackSlot.getNode())
11356 // Load the result.
11357 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
11358 FIST, StackSlot, MachinePointerInfo(),
11359 false, false, false, 0);
11361 // The node is the result.
11365 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
11367 MVT VT = Op.getSimpleValueType();
11368 SDValue In = Op.getOperand(0);
11369 MVT SVT = In.getSimpleValueType();
11371 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
11373 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
11374 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
11375 In, DAG.getUNDEF(SVT)));
11378 static SDValue LowerFABS(SDValue Op, SelectionDAG &DAG) {
11379 LLVMContext *Context = DAG.getContext();
11381 MVT VT = Op.getSimpleValueType();
11383 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
11384 if (VT.isVector()) {
11385 EltVT = VT.getVectorElementType();
11386 NumElts = VT.getVectorNumElements();
11389 if (EltVT == MVT::f64)
11390 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11391 APInt(64, ~(1ULL << 63))));
11393 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
11394 APInt(32, ~(1U << 31))));
11395 C = ConstantVector::getSplat(NumElts, C);
11396 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11397 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
11398 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11399 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
11400 MachinePointerInfo::getConstantPool(),
11401 false, false, false, Alignment);
11402 if (VT.isVector()) {
11403 MVT ANDVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
11404 return DAG.getNode(ISD::BITCAST, dl, VT,
11405 DAG.getNode(ISD::AND, dl, ANDVT,
11406 DAG.getNode(ISD::BITCAST, dl, ANDVT,
11408 DAG.getNode(ISD::BITCAST, dl, ANDVT, Mask)));
11410 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
11413 static SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG) {
11414 LLVMContext *Context = DAG.getContext();
11416 MVT VT = Op.getSimpleValueType();
11418 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
11419 if (VT.isVector()) {
11420 EltVT = VT.getVectorElementType();
11421 NumElts = VT.getVectorNumElements();
11424 if (EltVT == MVT::f64)
11425 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11426 APInt(64, 1ULL << 63)));
11428 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
11429 APInt(32, 1U << 31)));
11430 C = ConstantVector::getSplat(NumElts, C);
11431 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11432 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
11433 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11434 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
11435 MachinePointerInfo::getConstantPool(),
11436 false, false, false, Alignment);
11437 if (VT.isVector()) {
11438 MVT XORVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits()/64);
11439 return DAG.getNode(ISD::BITCAST, dl, VT,
11440 DAG.getNode(ISD::XOR, dl, XORVT,
11441 DAG.getNode(ISD::BITCAST, dl, XORVT,
11443 DAG.getNode(ISD::BITCAST, dl, XORVT, Mask)));
11446 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
11449 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
11450 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11451 LLVMContext *Context = DAG.getContext();
11452 SDValue Op0 = Op.getOperand(0);
11453 SDValue Op1 = Op.getOperand(1);
11455 MVT VT = Op.getSimpleValueType();
11456 MVT SrcVT = Op1.getSimpleValueType();
11458 // If second operand is smaller, extend it first.
11459 if (SrcVT.bitsLT(VT)) {
11460 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
11463 // And if it is bigger, shrink it first.
11464 if (SrcVT.bitsGT(VT)) {
11465 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
11469 // At this point the operands and the result should have the same
11470 // type, and that won't be f80 since that is not custom lowered.
11472 // First get the sign bit of second operand.
11473 SmallVector<Constant*,4> CV;
11474 if (SrcVT == MVT::f64) {
11475 const fltSemantics &Sem = APFloat::IEEEdouble;
11476 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 1ULL << 63))));
11477 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
11479 const fltSemantics &Sem = APFloat::IEEEsingle;
11480 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 1U << 31))));
11481 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
11482 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
11483 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
11485 Constant *C = ConstantVector::get(CV);
11486 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
11487 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
11488 MachinePointerInfo::getConstantPool(),
11489 false, false, false, 16);
11490 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
11492 // Shift sign bit right or left if the two operands have different types.
11493 if (SrcVT.bitsGT(VT)) {
11494 // Op0 is MVT::f32, Op1 is MVT::f64.
11495 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
11496 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
11497 DAG.getConstant(32, MVT::i32));
11498 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
11499 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
11500 DAG.getIntPtrConstant(0));
11503 // Clear first operand sign bit.
11505 if (VT == MVT::f64) {
11506 const fltSemantics &Sem = APFloat::IEEEdouble;
11507 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
11508 APInt(64, ~(1ULL << 63)))));
11509 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
11511 const fltSemantics &Sem = APFloat::IEEEsingle;
11512 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
11513 APInt(32, ~(1U << 31)))));
11514 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
11515 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
11516 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
11518 C = ConstantVector::get(CV);
11519 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
11520 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
11521 MachinePointerInfo::getConstantPool(),
11522 false, false, false, 16);
11523 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
11525 // Or the value with the sign bit.
11526 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
11529 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
11530 SDValue N0 = Op.getOperand(0);
11532 MVT VT = Op.getSimpleValueType();
11534 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
11535 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
11536 DAG.getConstant(1, VT));
11537 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
11540 // LowerVectorAllZeroTest - Check whether an OR'd tree is PTEST-able.
11542 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
11543 SelectionDAG &DAG) {
11544 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
11546 if (!Subtarget->hasSSE41())
11549 if (!Op->hasOneUse())
11552 SDNode *N = Op.getNode();
11555 SmallVector<SDValue, 8> Opnds;
11556 DenseMap<SDValue, unsigned> VecInMap;
11557 SmallVector<SDValue, 8> VecIns;
11558 EVT VT = MVT::Other;
11560 // Recognize a special case where a vector is casted into wide integer to
11562 Opnds.push_back(N->getOperand(0));
11563 Opnds.push_back(N->getOperand(1));
11565 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
11566 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
11567 // BFS traverse all OR'd operands.
11568 if (I->getOpcode() == ISD::OR) {
11569 Opnds.push_back(I->getOperand(0));
11570 Opnds.push_back(I->getOperand(1));
11571 // Re-evaluate the number of nodes to be traversed.
11572 e += 2; // 2 more nodes (LHS and RHS) are pushed.
11576 // Quit if a non-EXTRACT_VECTOR_ELT
11577 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11580 // Quit if without a constant index.
11581 SDValue Idx = I->getOperand(1);
11582 if (!isa<ConstantSDNode>(Idx))
11585 SDValue ExtractedFromVec = I->getOperand(0);
11586 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
11587 if (M == VecInMap.end()) {
11588 VT = ExtractedFromVec.getValueType();
11589 // Quit if not 128/256-bit vector.
11590 if (!VT.is128BitVector() && !VT.is256BitVector())
11592 // Quit if not the same type.
11593 if (VecInMap.begin() != VecInMap.end() &&
11594 VT != VecInMap.begin()->first.getValueType())
11596 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
11597 VecIns.push_back(ExtractedFromVec);
11599 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
11602 assert((VT.is128BitVector() || VT.is256BitVector()) &&
11603 "Not extracted from 128-/256-bit vector.");
11605 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
11607 for (DenseMap<SDValue, unsigned>::const_iterator
11608 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
11609 // Quit if not all elements are used.
11610 if (I->second != FullMask)
11614 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
11616 // Cast all vectors into TestVT for PTEST.
11617 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
11618 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
11620 // If more than one full vectors are evaluated, OR them first before PTEST.
11621 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
11622 // Each iteration will OR 2 nodes and append the result until there is only
11623 // 1 node left, i.e. the final OR'd value of all vectors.
11624 SDValue LHS = VecIns[Slot];
11625 SDValue RHS = VecIns[Slot + 1];
11626 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
11629 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
11630 VecIns.back(), VecIns.back());
11633 /// \brief return true if \c Op has a use that doesn't just read flags.
11634 static bool hasNonFlagsUse(SDValue Op) {
11635 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
11637 SDNode *User = *UI;
11638 unsigned UOpNo = UI.getOperandNo();
11639 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
11640 // Look pass truncate.
11641 UOpNo = User->use_begin().getOperandNo();
11642 User = *User->use_begin();
11645 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
11646 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
11652 /// Emit nodes that will be selected as "test Op0,Op0", or something
11654 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
11655 SelectionDAG &DAG) const {
11656 if (Op.getValueType() == MVT::i1)
11657 // KORTEST instruction should be selected
11658 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
11659 DAG.getConstant(0, Op.getValueType()));
11661 // CF and OF aren't always set the way we want. Determine which
11662 // of these we need.
11663 bool NeedCF = false;
11664 bool NeedOF = false;
11667 case X86::COND_A: case X86::COND_AE:
11668 case X86::COND_B: case X86::COND_BE:
11671 case X86::COND_G: case X86::COND_GE:
11672 case X86::COND_L: case X86::COND_LE:
11673 case X86::COND_O: case X86::COND_NO: {
11674 // Check if we really need to set the
11675 // Overflow flag. If NoSignedWrap is present
11676 // that is not actually needed.
11677 switch (Op->getOpcode()) {
11682 const BinaryWithFlagsSDNode *BinNode =
11683 cast<BinaryWithFlagsSDNode>(Op.getNode());
11684 if (BinNode->hasNoSignedWrap())
11694 // See if we can use the EFLAGS value from the operand instead of
11695 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
11696 // we prove that the arithmetic won't overflow, we can't use OF or CF.
11697 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
11698 // Emit a CMP with 0, which is the TEST pattern.
11699 //if (Op.getValueType() == MVT::i1)
11700 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
11701 // DAG.getConstant(0, MVT::i1));
11702 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
11703 DAG.getConstant(0, Op.getValueType()));
11705 unsigned Opcode = 0;
11706 unsigned NumOperands = 0;
11708 // Truncate operations may prevent the merge of the SETCC instruction
11709 // and the arithmetic instruction before it. Attempt to truncate the operands
11710 // of the arithmetic instruction and use a reduced bit-width instruction.
11711 bool NeedTruncation = false;
11712 SDValue ArithOp = Op;
11713 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
11714 SDValue Arith = Op->getOperand(0);
11715 // Both the trunc and the arithmetic op need to have one user each.
11716 if (Arith->hasOneUse())
11717 switch (Arith.getOpcode()) {
11724 NeedTruncation = true;
11730 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
11731 // which may be the result of a CAST. We use the variable 'Op', which is the
11732 // non-casted variable when we check for possible users.
11733 switch (ArithOp.getOpcode()) {
11735 // Due to an isel shortcoming, be conservative if this add is likely to be
11736 // selected as part of a load-modify-store instruction. When the root node
11737 // in a match is a store, isel doesn't know how to remap non-chain non-flag
11738 // uses of other nodes in the match, such as the ADD in this case. This
11739 // leads to the ADD being left around and reselected, with the result being
11740 // two adds in the output. Alas, even if none our users are stores, that
11741 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
11742 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
11743 // climbing the DAG back to the root, and it doesn't seem to be worth the
11745 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
11746 UE = Op.getNode()->use_end(); UI != UE; ++UI)
11747 if (UI->getOpcode() != ISD::CopyToReg &&
11748 UI->getOpcode() != ISD::SETCC &&
11749 UI->getOpcode() != ISD::STORE)
11752 if (ConstantSDNode *C =
11753 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
11754 // An add of one will be selected as an INC.
11755 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
11756 Opcode = X86ISD::INC;
11761 // An add of negative one (subtract of one) will be selected as a DEC.
11762 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
11763 Opcode = X86ISD::DEC;
11769 // Otherwise use a regular EFLAGS-setting add.
11770 Opcode = X86ISD::ADD;
11775 // If we have a constant logical shift that's only used in a comparison
11776 // against zero turn it into an equivalent AND. This allows turning it into
11777 // a TEST instruction later.
11778 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
11779 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
11780 EVT VT = Op.getValueType();
11781 unsigned BitWidth = VT.getSizeInBits();
11782 unsigned ShAmt = Op->getConstantOperandVal(1);
11783 if (ShAmt >= BitWidth) // Avoid undefined shifts.
11785 APInt Mask = ArithOp.getOpcode() == ISD::SRL
11786 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
11787 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
11788 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
11790 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
11791 DAG.getConstant(Mask, VT));
11792 DAG.ReplaceAllUsesWith(Op, New);
11798 // If the primary and result isn't used, don't bother using X86ISD::AND,
11799 // because a TEST instruction will be better.
11800 if (!hasNonFlagsUse(Op))
11806 // Due to the ISEL shortcoming noted above, be conservative if this op is
11807 // likely to be selected as part of a load-modify-store instruction.
11808 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
11809 UE = Op.getNode()->use_end(); UI != UE; ++UI)
11810 if (UI->getOpcode() == ISD::STORE)
11813 // Otherwise use a regular EFLAGS-setting instruction.
11814 switch (ArithOp.getOpcode()) {
11815 default: llvm_unreachable("unexpected operator!");
11816 case ISD::SUB: Opcode = X86ISD::SUB; break;
11817 case ISD::XOR: Opcode = X86ISD::XOR; break;
11818 case ISD::AND: Opcode = X86ISD::AND; break;
11820 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
11821 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
11822 if (EFLAGS.getNode())
11825 Opcode = X86ISD::OR;
11839 return SDValue(Op.getNode(), 1);
11845 // If we found that truncation is beneficial, perform the truncation and
11847 if (NeedTruncation) {
11848 EVT VT = Op.getValueType();
11849 SDValue WideVal = Op->getOperand(0);
11850 EVT WideVT = WideVal.getValueType();
11851 unsigned ConvertedOp = 0;
11852 // Use a target machine opcode to prevent further DAGCombine
11853 // optimizations that may separate the arithmetic operations
11854 // from the setcc node.
11855 switch (WideVal.getOpcode()) {
11857 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
11858 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
11859 case ISD::AND: ConvertedOp = X86ISD::AND; break;
11860 case ISD::OR: ConvertedOp = X86ISD::OR; break;
11861 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
11865 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11866 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
11867 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
11868 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
11869 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
11875 // Emit a CMP with 0, which is the TEST pattern.
11876 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
11877 DAG.getConstant(0, Op.getValueType()));
11879 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
11880 SmallVector<SDValue, 4> Ops;
11881 for (unsigned i = 0; i != NumOperands; ++i)
11882 Ops.push_back(Op.getOperand(i));
11884 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
11885 DAG.ReplaceAllUsesWith(Op, New);
11886 return SDValue(New.getNode(), 1);
11889 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
11891 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
11892 SDLoc dl, SelectionDAG &DAG) const {
11893 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
11894 if (C->getAPIntValue() == 0)
11895 return EmitTest(Op0, X86CC, dl, DAG);
11897 if (Op0.getValueType() == MVT::i1)
11898 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
11901 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
11902 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
11903 // Do the comparison at i32 if it's smaller, besides the Atom case.
11904 // This avoids subregister aliasing issues. Keep the smaller reference
11905 // if we're optimizing for size, however, as that'll allow better folding
11906 // of memory operations.
11907 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
11908 !DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
11909 AttributeSet::FunctionIndex, Attribute::MinSize) &&
11910 !Subtarget->isAtom()) {
11911 unsigned ExtendOp =
11912 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
11913 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
11914 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
11916 // Use SUB instead of CMP to enable CSE between SUB and CMP.
11917 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
11918 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
11920 return SDValue(Sub.getNode(), 1);
11922 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
11925 /// Convert a comparison if required by the subtarget.
11926 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
11927 SelectionDAG &DAG) const {
11928 // If the subtarget does not support the FUCOMI instruction, floating-point
11929 // comparisons have to be converted.
11930 if (Subtarget->hasCMov() ||
11931 Cmp.getOpcode() != X86ISD::CMP ||
11932 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
11933 !Cmp.getOperand(1).getValueType().isFloatingPoint())
11936 // The instruction selector will select an FUCOM instruction instead of
11937 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
11938 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
11939 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
11941 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
11942 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
11943 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
11944 DAG.getConstant(8, MVT::i8));
11945 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
11946 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
11949 static bool isAllOnes(SDValue V) {
11950 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
11951 return C && C->isAllOnesValue();
11954 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
11955 /// if it's possible.
11956 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
11957 SDLoc dl, SelectionDAG &DAG) const {
11958 SDValue Op0 = And.getOperand(0);
11959 SDValue Op1 = And.getOperand(1);
11960 if (Op0.getOpcode() == ISD::TRUNCATE)
11961 Op0 = Op0.getOperand(0);
11962 if (Op1.getOpcode() == ISD::TRUNCATE)
11963 Op1 = Op1.getOperand(0);
11966 if (Op1.getOpcode() == ISD::SHL)
11967 std::swap(Op0, Op1);
11968 if (Op0.getOpcode() == ISD::SHL) {
11969 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
11970 if (And00C->getZExtValue() == 1) {
11971 // If we looked past a truncate, check that it's only truncating away
11973 unsigned BitWidth = Op0.getValueSizeInBits();
11974 unsigned AndBitWidth = And.getValueSizeInBits();
11975 if (BitWidth > AndBitWidth) {
11977 DAG.computeKnownBits(Op0, Zeros, Ones);
11978 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
11982 RHS = Op0.getOperand(1);
11984 } else if (Op1.getOpcode() == ISD::Constant) {
11985 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
11986 uint64_t AndRHSVal = AndRHS->getZExtValue();
11987 SDValue AndLHS = Op0;
11989 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
11990 LHS = AndLHS.getOperand(0);
11991 RHS = AndLHS.getOperand(1);
11994 // Use BT if the immediate can't be encoded in a TEST instruction.
11995 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
11997 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
12001 if (LHS.getNode()) {
12002 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
12003 // instruction. Since the shift amount is in-range-or-undefined, we know
12004 // that doing a bittest on the i32 value is ok. We extend to i32 because
12005 // the encoding for the i16 version is larger than the i32 version.
12006 // Also promote i16 to i32 for performance / code size reason.
12007 if (LHS.getValueType() == MVT::i8 ||
12008 LHS.getValueType() == MVT::i16)
12009 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
12011 // If the operand types disagree, extend the shift amount to match. Since
12012 // BT ignores high bits (like shifts) we can use anyextend.
12013 if (LHS.getValueType() != RHS.getValueType())
12014 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
12016 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
12017 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
12018 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
12019 DAG.getConstant(Cond, MVT::i8), BT);
12025 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
12027 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
12032 // SSE Condition code mapping:
12041 switch (SetCCOpcode) {
12042 default: llvm_unreachable("Unexpected SETCC condition");
12044 case ISD::SETEQ: SSECC = 0; break;
12046 case ISD::SETGT: Swap = true; // Fallthrough
12048 case ISD::SETOLT: SSECC = 1; break;
12050 case ISD::SETGE: Swap = true; // Fallthrough
12052 case ISD::SETOLE: SSECC = 2; break;
12053 case ISD::SETUO: SSECC = 3; break;
12055 case ISD::SETNE: SSECC = 4; break;
12056 case ISD::SETULE: Swap = true; // Fallthrough
12057 case ISD::SETUGE: SSECC = 5; break;
12058 case ISD::SETULT: Swap = true; // Fallthrough
12059 case ISD::SETUGT: SSECC = 6; break;
12060 case ISD::SETO: SSECC = 7; break;
12062 case ISD::SETONE: SSECC = 8; break;
12065 std::swap(Op0, Op1);
12070 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
12071 // ones, and then concatenate the result back.
12072 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
12073 MVT VT = Op.getSimpleValueType();
12075 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
12076 "Unsupported value type for operation");
12078 unsigned NumElems = VT.getVectorNumElements();
12080 SDValue CC = Op.getOperand(2);
12082 // Extract the LHS vectors
12083 SDValue LHS = Op.getOperand(0);
12084 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
12085 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
12087 // Extract the RHS vectors
12088 SDValue RHS = Op.getOperand(1);
12089 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
12090 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
12092 // Issue the operation on the smaller types and concatenate the result back
12093 MVT EltVT = VT.getVectorElementType();
12094 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12095 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
12096 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
12097 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
12100 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
12101 const X86Subtarget *Subtarget) {
12102 SDValue Op0 = Op.getOperand(0);
12103 SDValue Op1 = Op.getOperand(1);
12104 SDValue CC = Op.getOperand(2);
12105 MVT VT = Op.getSimpleValueType();
12108 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 32 &&
12109 Op.getValueType().getScalarType() == MVT::i1 &&
12110 "Cannot set masked compare for this operation");
12112 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
12114 bool Unsigned = false;
12117 switch (SetCCOpcode) {
12118 default: llvm_unreachable("Unexpected SETCC condition");
12119 case ISD::SETNE: SSECC = 4; break;
12120 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
12121 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
12122 case ISD::SETLT: Swap = true; //fall-through
12123 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
12124 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
12125 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
12126 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
12127 case ISD::SETULE: Unsigned = true; //fall-through
12128 case ISD::SETLE: SSECC = 2; break;
12132 std::swap(Op0, Op1);
12134 return DAG.getNode(Opc, dl, VT, Op0, Op1);
12135 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
12136 return DAG.getNode(Opc, dl, VT, Op0, Op1,
12137 DAG.getConstant(SSECC, MVT::i8));
12140 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
12141 /// operand \p Op1. If non-trivial (for example because it's not constant)
12142 /// return an empty value.
12143 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
12145 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
12149 MVT VT = Op1.getSimpleValueType();
12150 MVT EVT = VT.getVectorElementType();
12151 unsigned n = VT.getVectorNumElements();
12152 SmallVector<SDValue, 8> ULTOp1;
12154 for (unsigned i = 0; i < n; ++i) {
12155 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
12156 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
12159 // Avoid underflow.
12160 APInt Val = Elt->getAPIntValue();
12164 ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
12167 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
12170 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
12171 SelectionDAG &DAG) {
12172 SDValue Op0 = Op.getOperand(0);
12173 SDValue Op1 = Op.getOperand(1);
12174 SDValue CC = Op.getOperand(2);
12175 MVT VT = Op.getSimpleValueType();
12176 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
12177 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
12182 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
12183 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
12186 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
12187 unsigned Opc = X86ISD::CMPP;
12188 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
12189 assert(VT.getVectorNumElements() <= 16);
12190 Opc = X86ISD::CMPM;
12192 // In the two special cases we can't handle, emit two comparisons.
12195 unsigned CombineOpc;
12196 if (SetCCOpcode == ISD::SETUEQ) {
12197 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
12199 assert(SetCCOpcode == ISD::SETONE);
12200 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
12203 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
12204 DAG.getConstant(CC0, MVT::i8));
12205 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
12206 DAG.getConstant(CC1, MVT::i8));
12207 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
12209 // Handle all other FP comparisons here.
12210 return DAG.getNode(Opc, dl, VT, Op0, Op1,
12211 DAG.getConstant(SSECC, MVT::i8));
12214 // Break 256-bit integer vector compare into smaller ones.
12215 if (VT.is256BitVector() && !Subtarget->hasInt256())
12216 return Lower256IntVSETCC(Op, DAG);
12218 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
12219 EVT OpVT = Op1.getValueType();
12220 if (Subtarget->hasAVX512()) {
12221 if (Op1.getValueType().is512BitVector() ||
12222 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
12223 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
12225 // In AVX-512 architecture setcc returns mask with i1 elements,
12226 // But there is no compare instruction for i8 and i16 elements.
12227 // We are not talking about 512-bit operands in this case, these
12228 // types are illegal.
12230 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
12231 OpVT.getVectorElementType().getSizeInBits() >= 8))
12232 return DAG.getNode(ISD::TRUNCATE, dl, VT,
12233 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
12236 // We are handling one of the integer comparisons here. Since SSE only has
12237 // GT and EQ comparisons for integer, swapping operands and multiple
12238 // operations may be required for some comparisons.
12240 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
12241 bool Subus = false;
12243 switch (SetCCOpcode) {
12244 default: llvm_unreachable("Unexpected SETCC condition");
12245 case ISD::SETNE: Invert = true;
12246 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
12247 case ISD::SETLT: Swap = true;
12248 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
12249 case ISD::SETGE: Swap = true;
12250 case ISD::SETLE: Opc = X86ISD::PCMPGT;
12251 Invert = true; break;
12252 case ISD::SETULT: Swap = true;
12253 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
12254 FlipSigns = true; break;
12255 case ISD::SETUGE: Swap = true;
12256 case ISD::SETULE: Opc = X86ISD::PCMPGT;
12257 FlipSigns = true; Invert = true; break;
12260 // Special case: Use min/max operations for SETULE/SETUGE
12261 MVT VET = VT.getVectorElementType();
12263 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
12264 || (Subtarget->hasSSE2() && (VET == MVT::i8));
12267 switch (SetCCOpcode) {
12269 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
12270 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
12273 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
12276 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
12277 if (!MinMax && hasSubus) {
12278 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
12280 // t = psubus Op0, Op1
12281 // pcmpeq t, <0..0>
12282 switch (SetCCOpcode) {
12284 case ISD::SETULT: {
12285 // If the comparison is against a constant we can turn this into a
12286 // setule. With psubus, setule does not require a swap. This is
12287 // beneficial because the constant in the register is no longer
12288 // destructed as the destination so it can be hoisted out of a loop.
12289 // Only do this pre-AVX since vpcmp* is no longer destructive.
12290 if (Subtarget->hasAVX())
12292 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
12293 if (ULEOp1.getNode()) {
12295 Subus = true; Invert = false; Swap = false;
12299 // Psubus is better than flip-sign because it requires no inversion.
12300 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
12301 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
12305 Opc = X86ISD::SUBUS;
12311 std::swap(Op0, Op1);
12313 // Check that the operation in question is available (most are plain SSE2,
12314 // but PCMPGTQ and PCMPEQQ have different requirements).
12315 if (VT == MVT::v2i64) {
12316 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
12317 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
12319 // First cast everything to the right type.
12320 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
12321 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
12323 // Since SSE has no unsigned integer comparisons, we need to flip the sign
12324 // bits of the inputs before performing those operations. The lower
12325 // compare is always unsigned.
12328 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
12330 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
12331 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
12332 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
12333 Sign, Zero, Sign, Zero);
12335 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
12336 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
12338 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
12339 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
12340 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
12342 // Create masks for only the low parts/high parts of the 64 bit integers.
12343 static const int MaskHi[] = { 1, 1, 3, 3 };
12344 static const int MaskLo[] = { 0, 0, 2, 2 };
12345 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
12346 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
12347 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
12349 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
12350 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
12353 Result = DAG.getNOT(dl, Result, MVT::v4i32);
12355 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
12358 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
12359 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
12360 // pcmpeqd + pshufd + pand.
12361 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
12363 // First cast everything to the right type.
12364 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
12365 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
12368 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
12370 // Make sure the lower and upper halves are both all-ones.
12371 static const int Mask[] = { 1, 0, 3, 2 };
12372 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
12373 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
12376 Result = DAG.getNOT(dl, Result, MVT::v4i32);
12378 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
12382 // Since SSE has no unsigned integer comparisons, we need to flip the sign
12383 // bits of the inputs before performing those operations.
12385 EVT EltVT = VT.getVectorElementType();
12386 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
12387 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
12388 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
12391 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
12393 // If the logical-not of the result is required, perform that now.
12395 Result = DAG.getNOT(dl, Result, VT);
12398 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
12401 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
12402 getZeroVector(VT, Subtarget, DAG, dl));
12407 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
12409 MVT VT = Op.getSimpleValueType();
12411 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
12413 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
12414 && "SetCC type must be 8-bit or 1-bit integer");
12415 SDValue Op0 = Op.getOperand(0);
12416 SDValue Op1 = Op.getOperand(1);
12418 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
12420 // Optimize to BT if possible.
12421 // Lower (X & (1 << N)) == 0 to BT(X, N).
12422 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
12423 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
12424 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
12425 Op1.getOpcode() == ISD::Constant &&
12426 cast<ConstantSDNode>(Op1)->isNullValue() &&
12427 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
12428 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
12429 if (NewSetCC.getNode())
12433 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
12435 if (Op1.getOpcode() == ISD::Constant &&
12436 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
12437 cast<ConstantSDNode>(Op1)->isNullValue()) &&
12438 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
12440 // If the input is a setcc, then reuse the input setcc or use a new one with
12441 // the inverted condition.
12442 if (Op0.getOpcode() == X86ISD::SETCC) {
12443 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
12444 bool Invert = (CC == ISD::SETNE) ^
12445 cast<ConstantSDNode>(Op1)->isNullValue();
12449 CCode = X86::GetOppositeBranchCondition(CCode);
12450 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
12451 DAG.getConstant(CCode, MVT::i8),
12452 Op0.getOperand(1));
12454 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
12458 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
12459 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
12460 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
12462 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
12463 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
12466 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
12467 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
12468 if (X86CC == X86::COND_INVALID)
12471 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
12472 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
12473 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
12474 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
12476 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
12480 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
12481 static bool isX86LogicalCmp(SDValue Op) {
12482 unsigned Opc = Op.getNode()->getOpcode();
12483 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
12484 Opc == X86ISD::SAHF)
12486 if (Op.getResNo() == 1 &&
12487 (Opc == X86ISD::ADD ||
12488 Opc == X86ISD::SUB ||
12489 Opc == X86ISD::ADC ||
12490 Opc == X86ISD::SBB ||
12491 Opc == X86ISD::SMUL ||
12492 Opc == X86ISD::UMUL ||
12493 Opc == X86ISD::INC ||
12494 Opc == X86ISD::DEC ||
12495 Opc == X86ISD::OR ||
12496 Opc == X86ISD::XOR ||
12497 Opc == X86ISD::AND))
12500 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
12506 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
12507 if (V.getOpcode() != ISD::TRUNCATE)
12510 SDValue VOp0 = V.getOperand(0);
12511 unsigned InBits = VOp0.getValueSizeInBits();
12512 unsigned Bits = V.getValueSizeInBits();
12513 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
12516 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
12517 bool addTest = true;
12518 SDValue Cond = Op.getOperand(0);
12519 SDValue Op1 = Op.getOperand(1);
12520 SDValue Op2 = Op.getOperand(2);
12522 EVT VT = Op1.getValueType();
12525 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
12526 // are available. Otherwise fp cmovs get lowered into a less efficient branch
12527 // sequence later on.
12528 if (Cond.getOpcode() == ISD::SETCC &&
12529 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
12530 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
12531 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
12532 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
12533 int SSECC = translateX86FSETCC(
12534 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
12537 if (Subtarget->hasAVX512()) {
12538 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
12539 DAG.getConstant(SSECC, MVT::i8));
12540 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
12542 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
12543 DAG.getConstant(SSECC, MVT::i8));
12544 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
12545 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
12546 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
12550 if (Cond.getOpcode() == ISD::SETCC) {
12551 SDValue NewCond = LowerSETCC(Cond, DAG);
12552 if (NewCond.getNode())
12556 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
12557 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
12558 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
12559 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
12560 if (Cond.getOpcode() == X86ISD::SETCC &&
12561 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
12562 isZero(Cond.getOperand(1).getOperand(1))) {
12563 SDValue Cmp = Cond.getOperand(1);
12565 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
12567 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
12568 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
12569 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
12571 SDValue CmpOp0 = Cmp.getOperand(0);
12572 // Apply further optimizations for special cases
12573 // (select (x != 0), -1, 0) -> neg & sbb
12574 // (select (x == 0), 0, -1) -> neg & sbb
12575 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
12576 if (YC->isNullValue() &&
12577 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
12578 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
12579 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
12580 DAG.getConstant(0, CmpOp0.getValueType()),
12582 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
12583 DAG.getConstant(X86::COND_B, MVT::i8),
12584 SDValue(Neg.getNode(), 1));
12588 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
12589 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
12590 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
12592 SDValue Res = // Res = 0 or -1.
12593 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
12594 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
12596 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
12597 Res = DAG.getNOT(DL, Res, Res.getValueType());
12599 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
12600 if (!N2C || !N2C->isNullValue())
12601 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
12606 // Look past (and (setcc_carry (cmp ...)), 1).
12607 if (Cond.getOpcode() == ISD::AND &&
12608 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
12609 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
12610 if (C && C->getAPIntValue() == 1)
12611 Cond = Cond.getOperand(0);
12614 // If condition flag is set by a X86ISD::CMP, then use it as the condition
12615 // setting operand in place of the X86ISD::SETCC.
12616 unsigned CondOpcode = Cond.getOpcode();
12617 if (CondOpcode == X86ISD::SETCC ||
12618 CondOpcode == X86ISD::SETCC_CARRY) {
12619 CC = Cond.getOperand(0);
12621 SDValue Cmp = Cond.getOperand(1);
12622 unsigned Opc = Cmp.getOpcode();
12623 MVT VT = Op.getSimpleValueType();
12625 bool IllegalFPCMov = false;
12626 if (VT.isFloatingPoint() && !VT.isVector() &&
12627 !isScalarFPTypeInSSEReg(VT)) // FPStack?
12628 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
12630 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
12631 Opc == X86ISD::BT) { // FIXME
12635 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
12636 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
12637 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
12638 Cond.getOperand(0).getValueType() != MVT::i8)) {
12639 SDValue LHS = Cond.getOperand(0);
12640 SDValue RHS = Cond.getOperand(1);
12641 unsigned X86Opcode;
12644 switch (CondOpcode) {
12645 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
12646 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
12647 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
12648 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
12649 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
12650 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
12651 default: llvm_unreachable("unexpected overflowing operator");
12653 if (CondOpcode == ISD::UMULO)
12654 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
12657 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
12659 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
12661 if (CondOpcode == ISD::UMULO)
12662 Cond = X86Op.getValue(2);
12664 Cond = X86Op.getValue(1);
12666 CC = DAG.getConstant(X86Cond, MVT::i8);
12671 // Look pass the truncate if the high bits are known zero.
12672 if (isTruncWithZeroHighBitsInput(Cond, DAG))
12673 Cond = Cond.getOperand(0);
12675 // We know the result of AND is compared against zero. Try to match
12677 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
12678 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
12679 if (NewSetCC.getNode()) {
12680 CC = NewSetCC.getOperand(0);
12681 Cond = NewSetCC.getOperand(1);
12688 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
12689 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
12692 // a < b ? -1 : 0 -> RES = ~setcc_carry
12693 // a < b ? 0 : -1 -> RES = setcc_carry
12694 // a >= b ? -1 : 0 -> RES = setcc_carry
12695 // a >= b ? 0 : -1 -> RES = ~setcc_carry
12696 if (Cond.getOpcode() == X86ISD::SUB) {
12697 Cond = ConvertCmpIfNecessary(Cond, DAG);
12698 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
12700 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
12701 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
12702 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
12703 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
12704 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
12705 return DAG.getNOT(DL, Res, Res.getValueType());
12710 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
12711 // widen the cmov and push the truncate through. This avoids introducing a new
12712 // branch during isel and doesn't add any extensions.
12713 if (Op.getValueType() == MVT::i8 &&
12714 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
12715 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
12716 if (T1.getValueType() == T2.getValueType() &&
12717 // Blacklist CopyFromReg to avoid partial register stalls.
12718 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
12719 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
12720 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
12721 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
12725 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
12726 // condition is true.
12727 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
12728 SDValue Ops[] = { Op2, Op1, CC, Cond };
12729 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
12732 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, SelectionDAG &DAG) {
12733 MVT VT = Op->getSimpleValueType(0);
12734 SDValue In = Op->getOperand(0);
12735 MVT InVT = In.getSimpleValueType();
12738 unsigned int NumElts = VT.getVectorNumElements();
12739 if (NumElts != 8 && NumElts != 16)
12742 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12743 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
12745 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12746 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12748 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
12749 Constant *C = ConstantInt::get(*DAG.getContext(),
12750 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
12752 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
12753 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
12754 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
12755 MachinePointerInfo::getConstantPool(),
12756 false, false, false, Alignment);
12757 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
12758 if (VT.is512BitVector())
12760 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
12763 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12764 SelectionDAG &DAG) {
12765 MVT VT = Op->getSimpleValueType(0);
12766 SDValue In = Op->getOperand(0);
12767 MVT InVT = In.getSimpleValueType();
12770 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
12771 return LowerSIGN_EXTEND_AVX512(Op, DAG);
12773 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
12774 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
12775 (VT != MVT::v16i16 || InVT != MVT::v16i8))
12778 if (Subtarget->hasInt256())
12779 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
12781 // Optimize vectors in AVX mode
12782 // Sign extend v8i16 to v8i32 and
12785 // Divide input vector into two parts
12786 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
12787 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
12788 // concat the vectors to original VT
12790 unsigned NumElems = InVT.getVectorNumElements();
12791 SDValue Undef = DAG.getUNDEF(InVT);
12793 SmallVector<int,8> ShufMask1(NumElems, -1);
12794 for (unsigned i = 0; i != NumElems/2; ++i)
12797 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
12799 SmallVector<int,8> ShufMask2(NumElems, -1);
12800 for (unsigned i = 0; i != NumElems/2; ++i)
12801 ShufMask2[i] = i + NumElems/2;
12803 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
12805 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
12806 VT.getVectorNumElements()/2);
12808 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
12809 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
12811 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12814 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
12815 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
12816 // from the AND / OR.
12817 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
12818 Opc = Op.getOpcode();
12819 if (Opc != ISD::OR && Opc != ISD::AND)
12821 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
12822 Op.getOperand(0).hasOneUse() &&
12823 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
12824 Op.getOperand(1).hasOneUse());
12827 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
12828 // 1 and that the SETCC node has a single use.
12829 static bool isXor1OfSetCC(SDValue Op) {
12830 if (Op.getOpcode() != ISD::XOR)
12832 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
12833 if (N1C && N1C->getAPIntValue() == 1) {
12834 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
12835 Op.getOperand(0).hasOneUse();
12840 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
12841 bool addTest = true;
12842 SDValue Chain = Op.getOperand(0);
12843 SDValue Cond = Op.getOperand(1);
12844 SDValue Dest = Op.getOperand(2);
12847 bool Inverted = false;
12849 if (Cond.getOpcode() == ISD::SETCC) {
12850 // Check for setcc([su]{add,sub,mul}o == 0).
12851 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
12852 isa<ConstantSDNode>(Cond.getOperand(1)) &&
12853 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
12854 Cond.getOperand(0).getResNo() == 1 &&
12855 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
12856 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
12857 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
12858 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
12859 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
12860 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
12862 Cond = Cond.getOperand(0);
12864 SDValue NewCond = LowerSETCC(Cond, DAG);
12865 if (NewCond.getNode())
12870 // FIXME: LowerXALUO doesn't handle these!!
12871 else if (Cond.getOpcode() == X86ISD::ADD ||
12872 Cond.getOpcode() == X86ISD::SUB ||
12873 Cond.getOpcode() == X86ISD::SMUL ||
12874 Cond.getOpcode() == X86ISD::UMUL)
12875 Cond = LowerXALUO(Cond, DAG);
12878 // Look pass (and (setcc_carry (cmp ...)), 1).
12879 if (Cond.getOpcode() == ISD::AND &&
12880 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
12881 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
12882 if (C && C->getAPIntValue() == 1)
12883 Cond = Cond.getOperand(0);
12886 // If condition flag is set by a X86ISD::CMP, then use it as the condition
12887 // setting operand in place of the X86ISD::SETCC.
12888 unsigned CondOpcode = Cond.getOpcode();
12889 if (CondOpcode == X86ISD::SETCC ||
12890 CondOpcode == X86ISD::SETCC_CARRY) {
12891 CC = Cond.getOperand(0);
12893 SDValue Cmp = Cond.getOperand(1);
12894 unsigned Opc = Cmp.getOpcode();
12895 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
12896 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
12900 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
12904 // These can only come from an arithmetic instruction with overflow,
12905 // e.g. SADDO, UADDO.
12906 Cond = Cond.getNode()->getOperand(1);
12912 CondOpcode = Cond.getOpcode();
12913 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
12914 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
12915 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
12916 Cond.getOperand(0).getValueType() != MVT::i8)) {
12917 SDValue LHS = Cond.getOperand(0);
12918 SDValue RHS = Cond.getOperand(1);
12919 unsigned X86Opcode;
12922 // Keep this in sync with LowerXALUO, otherwise we might create redundant
12923 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
12925 switch (CondOpcode) {
12926 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
12928 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
12930 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
12933 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
12934 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
12936 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
12938 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
12941 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
12942 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
12943 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
12944 default: llvm_unreachable("unexpected overflowing operator");
12947 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
12948 if (CondOpcode == ISD::UMULO)
12949 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
12952 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
12954 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
12956 if (CondOpcode == ISD::UMULO)
12957 Cond = X86Op.getValue(2);
12959 Cond = X86Op.getValue(1);
12961 CC = DAG.getConstant(X86Cond, MVT::i8);
12965 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
12966 SDValue Cmp = Cond.getOperand(0).getOperand(1);
12967 if (CondOpc == ISD::OR) {
12968 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
12969 // two branches instead of an explicit OR instruction with a
12971 if (Cmp == Cond.getOperand(1).getOperand(1) &&
12972 isX86LogicalCmp(Cmp)) {
12973 CC = Cond.getOperand(0).getOperand(0);
12974 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
12975 Chain, Dest, CC, Cmp);
12976 CC = Cond.getOperand(1).getOperand(0);
12980 } else { // ISD::AND
12981 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
12982 // two branches instead of an explicit AND instruction with a
12983 // separate test. However, we only do this if this block doesn't
12984 // have a fall-through edge, because this requires an explicit
12985 // jmp when the condition is false.
12986 if (Cmp == Cond.getOperand(1).getOperand(1) &&
12987 isX86LogicalCmp(Cmp) &&
12988 Op.getNode()->hasOneUse()) {
12989 X86::CondCode CCode =
12990 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
12991 CCode = X86::GetOppositeBranchCondition(CCode);
12992 CC = DAG.getConstant(CCode, MVT::i8);
12993 SDNode *User = *Op.getNode()->use_begin();
12994 // Look for an unconditional branch following this conditional branch.
12995 // We need this because we need to reverse the successors in order
12996 // to implement FCMP_OEQ.
12997 if (User->getOpcode() == ISD::BR) {
12998 SDValue FalseBB = User->getOperand(1);
13000 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13001 assert(NewBR == User);
13005 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13006 Chain, Dest, CC, Cmp);
13007 X86::CondCode CCode =
13008 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
13009 CCode = X86::GetOppositeBranchCondition(CCode);
13010 CC = DAG.getConstant(CCode, MVT::i8);
13016 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
13017 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
13018 // It should be transformed during dag combiner except when the condition
13019 // is set by a arithmetics with overflow node.
13020 X86::CondCode CCode =
13021 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
13022 CCode = X86::GetOppositeBranchCondition(CCode);
13023 CC = DAG.getConstant(CCode, MVT::i8);
13024 Cond = Cond.getOperand(0).getOperand(1);
13026 } else if (Cond.getOpcode() == ISD::SETCC &&
13027 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
13028 // For FCMP_OEQ, we can emit
13029 // two branches instead of an explicit AND instruction with a
13030 // separate test. However, we only do this if this block doesn't
13031 // have a fall-through edge, because this requires an explicit
13032 // jmp when the condition is false.
13033 if (Op.getNode()->hasOneUse()) {
13034 SDNode *User = *Op.getNode()->use_begin();
13035 // Look for an unconditional branch following this conditional branch.
13036 // We need this because we need to reverse the successors in order
13037 // to implement FCMP_OEQ.
13038 if (User->getOpcode() == ISD::BR) {
13039 SDValue FalseBB = User->getOperand(1);
13041 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13042 assert(NewBR == User);
13046 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13047 Cond.getOperand(0), Cond.getOperand(1));
13048 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13049 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13050 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13051 Chain, Dest, CC, Cmp);
13052 CC = DAG.getConstant(X86::COND_P, MVT::i8);
13057 } else if (Cond.getOpcode() == ISD::SETCC &&
13058 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
13059 // For FCMP_UNE, we can emit
13060 // two branches instead of an explicit AND instruction with a
13061 // separate test. However, we only do this if this block doesn't
13062 // have a fall-through edge, because this requires an explicit
13063 // jmp when the condition is false.
13064 if (Op.getNode()->hasOneUse()) {
13065 SDNode *User = *Op.getNode()->use_begin();
13066 // Look for an unconditional branch following this conditional branch.
13067 // We need this because we need to reverse the successors in order
13068 // to implement FCMP_UNE.
13069 if (User->getOpcode() == ISD::BR) {
13070 SDValue FalseBB = User->getOperand(1);
13072 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13073 assert(NewBR == User);
13076 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13077 Cond.getOperand(0), Cond.getOperand(1));
13078 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13079 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13080 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13081 Chain, Dest, CC, Cmp);
13082 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
13092 // Look pass the truncate if the high bits are known zero.
13093 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13094 Cond = Cond.getOperand(0);
13096 // We know the result of AND is compared against zero. Try to match
13098 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13099 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
13100 if (NewSetCC.getNode()) {
13101 CC = NewSetCC.getOperand(0);
13102 Cond = NewSetCC.getOperand(1);
13109 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
13110 CC = DAG.getConstant(X86Cond, MVT::i8);
13111 Cond = EmitTest(Cond, X86Cond, dl, DAG);
13113 Cond = ConvertCmpIfNecessary(Cond, DAG);
13114 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13115 Chain, Dest, CC, Cond);
13118 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
13119 // Calls to _alloca is needed to probe the stack when allocating more than 4k
13120 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
13121 // that the guard pages used by the OS virtual memory manager are allocated in
13122 // correct sequence.
13124 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
13125 SelectionDAG &DAG) const {
13126 MachineFunction &MF = DAG.getMachineFunction();
13127 bool SplitStack = MF.shouldSplitStack();
13128 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMacho()) ||
13133 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13134 SDNode* Node = Op.getNode();
13136 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
13137 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
13138 " not tell us which reg is the stack pointer!");
13139 EVT VT = Node->getValueType(0);
13140 SDValue Tmp1 = SDValue(Node, 0);
13141 SDValue Tmp2 = SDValue(Node, 1);
13142 SDValue Tmp3 = Node->getOperand(2);
13143 SDValue Chain = Tmp1.getOperand(0);
13145 // Chain the dynamic stack allocation so that it doesn't modify the stack
13146 // pointer when other instructions are using the stack.
13147 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
13150 SDValue Size = Tmp2.getOperand(1);
13151 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
13152 Chain = SP.getValue(1);
13153 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
13154 const TargetFrameLowering &TFI = *DAG.getTarget().getFrameLowering();
13155 unsigned StackAlign = TFI.getStackAlignment();
13156 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
13157 if (Align > StackAlign)
13158 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
13159 DAG.getConstant(-(uint64_t)Align, VT));
13160 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
13162 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
13163 DAG.getIntPtrConstant(0, true), SDValue(),
13166 SDValue Ops[2] = { Tmp1, Tmp2 };
13167 return DAG.getMergeValues(Ops, dl);
13171 SDValue Chain = Op.getOperand(0);
13172 SDValue Size = Op.getOperand(1);
13173 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
13174 EVT VT = Op.getNode()->getValueType(0);
13176 bool Is64Bit = Subtarget->is64Bit();
13177 EVT SPTy = Is64Bit ? MVT::i64 : MVT::i32;
13180 MachineRegisterInfo &MRI = MF.getRegInfo();
13183 // The 64 bit implementation of segmented stacks needs to clobber both r10
13184 // r11. This makes it impossible to use it along with nested parameters.
13185 const Function *F = MF.getFunction();
13187 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
13189 if (I->hasNestAttr())
13190 report_fatal_error("Cannot use segmented stacks with functions that "
13191 "have nested arguments.");
13194 const TargetRegisterClass *AddrRegClass =
13195 getRegClassFor(Subtarget->is64Bit() ? MVT::i64:MVT::i32);
13196 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
13197 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
13198 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
13199 DAG.getRegister(Vreg, SPTy));
13200 SDValue Ops1[2] = { Value, Chain };
13201 return DAG.getMergeValues(Ops1, dl);
13204 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
13206 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
13207 Flag = Chain.getValue(1);
13208 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
13210 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
13212 const X86RegisterInfo *RegInfo =
13213 static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
13214 unsigned SPReg = RegInfo->getStackRegister();
13215 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
13216 Chain = SP.getValue(1);
13219 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
13220 DAG.getConstant(-(uint64_t)Align, VT));
13221 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
13224 SDValue Ops1[2] = { SP, Chain };
13225 return DAG.getMergeValues(Ops1, dl);
13229 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
13230 MachineFunction &MF = DAG.getMachineFunction();
13231 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
13233 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
13236 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
13237 // vastart just stores the address of the VarArgsFrameIndex slot into the
13238 // memory location argument.
13239 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
13241 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
13242 MachinePointerInfo(SV), false, false, 0);
13246 // gp_offset (0 - 6 * 8)
13247 // fp_offset (48 - 48 + 8 * 16)
13248 // overflow_arg_area (point to parameters coming in memory).
13250 SmallVector<SDValue, 8> MemOps;
13251 SDValue FIN = Op.getOperand(1);
13253 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
13254 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
13256 FIN, MachinePointerInfo(SV), false, false, 0);
13257 MemOps.push_back(Store);
13260 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13261 FIN, DAG.getIntPtrConstant(4));
13262 Store = DAG.getStore(Op.getOperand(0), DL,
13263 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
13265 FIN, MachinePointerInfo(SV, 4), false, false, 0);
13266 MemOps.push_back(Store);
13268 // Store ptr to overflow_arg_area
13269 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13270 FIN, DAG.getIntPtrConstant(4));
13271 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
13273 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
13274 MachinePointerInfo(SV, 8),
13276 MemOps.push_back(Store);
13278 // Store ptr to reg_save_area.
13279 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13280 FIN, DAG.getIntPtrConstant(8));
13281 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
13283 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
13284 MachinePointerInfo(SV, 16), false, false, 0);
13285 MemOps.push_back(Store);
13286 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
13289 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
13290 assert(Subtarget->is64Bit() &&
13291 "LowerVAARG only handles 64-bit va_arg!");
13292 assert((Subtarget->isTargetLinux() ||
13293 Subtarget->isTargetDarwin()) &&
13294 "Unhandled target in LowerVAARG");
13295 assert(Op.getNode()->getNumOperands() == 4);
13296 SDValue Chain = Op.getOperand(0);
13297 SDValue SrcPtr = Op.getOperand(1);
13298 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
13299 unsigned Align = Op.getConstantOperandVal(3);
13302 EVT ArgVT = Op.getNode()->getValueType(0);
13303 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
13304 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
13307 // Decide which area this value should be read from.
13308 // TODO: Implement the AMD64 ABI in its entirety. This simple
13309 // selection mechanism works only for the basic types.
13310 if (ArgVT == MVT::f80) {
13311 llvm_unreachable("va_arg for f80 not yet implemented");
13312 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
13313 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
13314 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
13315 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
13317 llvm_unreachable("Unhandled argument type in LowerVAARG");
13320 if (ArgMode == 2) {
13321 // Sanity Check: Make sure using fp_offset makes sense.
13322 assert(!DAG.getTarget().Options.UseSoftFloat &&
13323 !(DAG.getMachineFunction()
13324 .getFunction()->getAttributes()
13325 .hasAttribute(AttributeSet::FunctionIndex,
13326 Attribute::NoImplicitFloat)) &&
13327 Subtarget->hasSSE1());
13330 // Insert VAARG_64 node into the DAG
13331 // VAARG_64 returns two values: Variable Argument Address, Chain
13332 SmallVector<SDValue, 11> InstOps;
13333 InstOps.push_back(Chain);
13334 InstOps.push_back(SrcPtr);
13335 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
13336 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
13337 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
13338 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
13339 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
13340 VTs, InstOps, MVT::i64,
13341 MachinePointerInfo(SV),
13343 /*Volatile=*/false,
13345 /*WriteMem=*/true);
13346 Chain = VAARG.getValue(1);
13348 // Load the next argument and return it
13349 return DAG.getLoad(ArgVT, dl,
13352 MachinePointerInfo(),
13353 false, false, false, 0);
13356 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
13357 SelectionDAG &DAG) {
13358 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
13359 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
13360 SDValue Chain = Op.getOperand(0);
13361 SDValue DstPtr = Op.getOperand(1);
13362 SDValue SrcPtr = Op.getOperand(2);
13363 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
13364 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
13367 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
13368 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
13370 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
13373 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
13374 // amount is a constant. Takes immediate version of shift as input.
13375 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
13376 SDValue SrcOp, uint64_t ShiftAmt,
13377 SelectionDAG &DAG) {
13378 MVT ElementType = VT.getVectorElementType();
13380 // Fold this packed shift into its first operand if ShiftAmt is 0.
13384 // Check for ShiftAmt >= element width
13385 if (ShiftAmt >= ElementType.getSizeInBits()) {
13386 if (Opc == X86ISD::VSRAI)
13387 ShiftAmt = ElementType.getSizeInBits() - 1;
13389 return DAG.getConstant(0, VT);
13392 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
13393 && "Unknown target vector shift-by-constant node");
13395 // Fold this packed vector shift into a build vector if SrcOp is a
13396 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
13397 if (VT == SrcOp.getSimpleValueType() &&
13398 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
13399 SmallVector<SDValue, 8> Elts;
13400 unsigned NumElts = SrcOp->getNumOperands();
13401 ConstantSDNode *ND;
13404 default: llvm_unreachable(nullptr);
13405 case X86ISD::VSHLI:
13406 for (unsigned i=0; i!=NumElts; ++i) {
13407 SDValue CurrentOp = SrcOp->getOperand(i);
13408 if (CurrentOp->getOpcode() == ISD::UNDEF) {
13409 Elts.push_back(CurrentOp);
13412 ND = cast<ConstantSDNode>(CurrentOp);
13413 const APInt &C = ND->getAPIntValue();
13414 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType));
13417 case X86ISD::VSRLI:
13418 for (unsigned i=0; i!=NumElts; ++i) {
13419 SDValue CurrentOp = SrcOp->getOperand(i);
13420 if (CurrentOp->getOpcode() == ISD::UNDEF) {
13421 Elts.push_back(CurrentOp);
13424 ND = cast<ConstantSDNode>(CurrentOp);
13425 const APInt &C = ND->getAPIntValue();
13426 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType));
13429 case X86ISD::VSRAI:
13430 for (unsigned i=0; i!=NumElts; ++i) {
13431 SDValue CurrentOp = SrcOp->getOperand(i);
13432 if (CurrentOp->getOpcode() == ISD::UNDEF) {
13433 Elts.push_back(CurrentOp);
13436 ND = cast<ConstantSDNode>(CurrentOp);
13437 const APInt &C = ND->getAPIntValue();
13438 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType));
13443 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
13446 return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8));
13449 // getTargetVShiftNode - Handle vector element shifts where the shift amount
13450 // may or may not be a constant. Takes immediate version of shift as input.
13451 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
13452 SDValue SrcOp, SDValue ShAmt,
13453 SelectionDAG &DAG) {
13454 assert(ShAmt.getValueType() == MVT::i32 && "ShAmt is not i32");
13456 // Catch shift-by-constant.
13457 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
13458 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
13459 CShAmt->getZExtValue(), DAG);
13461 // Change opcode to non-immediate version
13463 default: llvm_unreachable("Unknown target vector shift node");
13464 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
13465 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
13466 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
13469 // Need to build a vector containing shift amount
13470 // Shift amount is 32-bits, but SSE instructions read 64-bit, so fill with 0
13473 ShOps[1] = DAG.getConstant(0, MVT::i32);
13474 ShOps[2] = ShOps[3] = DAG.getUNDEF(MVT::i32);
13475 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, ShOps);
13477 // The return type has to be a 128-bit type with the same element
13478 // type as the input type.
13479 MVT EltVT = VT.getVectorElementType();
13480 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
13482 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
13483 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
13486 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
13488 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
13490 default: return SDValue(); // Don't custom lower most intrinsics.
13491 // Comparison intrinsics.
13492 case Intrinsic::x86_sse_comieq_ss:
13493 case Intrinsic::x86_sse_comilt_ss:
13494 case Intrinsic::x86_sse_comile_ss:
13495 case Intrinsic::x86_sse_comigt_ss:
13496 case Intrinsic::x86_sse_comige_ss:
13497 case Intrinsic::x86_sse_comineq_ss:
13498 case Intrinsic::x86_sse_ucomieq_ss:
13499 case Intrinsic::x86_sse_ucomilt_ss:
13500 case Intrinsic::x86_sse_ucomile_ss:
13501 case Intrinsic::x86_sse_ucomigt_ss:
13502 case Intrinsic::x86_sse_ucomige_ss:
13503 case Intrinsic::x86_sse_ucomineq_ss:
13504 case Intrinsic::x86_sse2_comieq_sd:
13505 case Intrinsic::x86_sse2_comilt_sd:
13506 case Intrinsic::x86_sse2_comile_sd:
13507 case Intrinsic::x86_sse2_comigt_sd:
13508 case Intrinsic::x86_sse2_comige_sd:
13509 case Intrinsic::x86_sse2_comineq_sd:
13510 case Intrinsic::x86_sse2_ucomieq_sd:
13511 case Intrinsic::x86_sse2_ucomilt_sd:
13512 case Intrinsic::x86_sse2_ucomile_sd:
13513 case Intrinsic::x86_sse2_ucomigt_sd:
13514 case Intrinsic::x86_sse2_ucomige_sd:
13515 case Intrinsic::x86_sse2_ucomineq_sd: {
13519 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
13520 case Intrinsic::x86_sse_comieq_ss:
13521 case Intrinsic::x86_sse2_comieq_sd:
13522 Opc = X86ISD::COMI;
13525 case Intrinsic::x86_sse_comilt_ss:
13526 case Intrinsic::x86_sse2_comilt_sd:
13527 Opc = X86ISD::COMI;
13530 case Intrinsic::x86_sse_comile_ss:
13531 case Intrinsic::x86_sse2_comile_sd:
13532 Opc = X86ISD::COMI;
13535 case Intrinsic::x86_sse_comigt_ss:
13536 case Intrinsic::x86_sse2_comigt_sd:
13537 Opc = X86ISD::COMI;
13540 case Intrinsic::x86_sse_comige_ss:
13541 case Intrinsic::x86_sse2_comige_sd:
13542 Opc = X86ISD::COMI;
13545 case Intrinsic::x86_sse_comineq_ss:
13546 case Intrinsic::x86_sse2_comineq_sd:
13547 Opc = X86ISD::COMI;
13550 case Intrinsic::x86_sse_ucomieq_ss:
13551 case Intrinsic::x86_sse2_ucomieq_sd:
13552 Opc = X86ISD::UCOMI;
13555 case Intrinsic::x86_sse_ucomilt_ss:
13556 case Intrinsic::x86_sse2_ucomilt_sd:
13557 Opc = X86ISD::UCOMI;
13560 case Intrinsic::x86_sse_ucomile_ss:
13561 case Intrinsic::x86_sse2_ucomile_sd:
13562 Opc = X86ISD::UCOMI;
13565 case Intrinsic::x86_sse_ucomigt_ss:
13566 case Intrinsic::x86_sse2_ucomigt_sd:
13567 Opc = X86ISD::UCOMI;
13570 case Intrinsic::x86_sse_ucomige_ss:
13571 case Intrinsic::x86_sse2_ucomige_sd:
13572 Opc = X86ISD::UCOMI;
13575 case Intrinsic::x86_sse_ucomineq_ss:
13576 case Intrinsic::x86_sse2_ucomineq_sd:
13577 Opc = X86ISD::UCOMI;
13582 SDValue LHS = Op.getOperand(1);
13583 SDValue RHS = Op.getOperand(2);
13584 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
13585 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
13586 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
13587 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13588 DAG.getConstant(X86CC, MVT::i8), Cond);
13589 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
13592 // Arithmetic intrinsics.
13593 case Intrinsic::x86_sse2_pmulu_dq:
13594 case Intrinsic::x86_avx2_pmulu_dq:
13595 return DAG.getNode(X86ISD::PMULUDQ, dl, Op.getValueType(),
13596 Op.getOperand(1), Op.getOperand(2));
13598 case Intrinsic::x86_sse41_pmuldq:
13599 case Intrinsic::x86_avx2_pmul_dq:
13600 return DAG.getNode(X86ISD::PMULDQ, dl, Op.getValueType(),
13601 Op.getOperand(1), Op.getOperand(2));
13603 case Intrinsic::x86_sse2_pmulhu_w:
13604 case Intrinsic::x86_avx2_pmulhu_w:
13605 return DAG.getNode(ISD::MULHU, dl, Op.getValueType(),
13606 Op.getOperand(1), Op.getOperand(2));
13608 case Intrinsic::x86_sse2_pmulh_w:
13609 case Intrinsic::x86_avx2_pmulh_w:
13610 return DAG.getNode(ISD::MULHS, dl, Op.getValueType(),
13611 Op.getOperand(1), Op.getOperand(2));
13613 // SSE2/AVX2 sub with unsigned saturation intrinsics
13614 case Intrinsic::x86_sse2_psubus_b:
13615 case Intrinsic::x86_sse2_psubus_w:
13616 case Intrinsic::x86_avx2_psubus_b:
13617 case Intrinsic::x86_avx2_psubus_w:
13618 return DAG.getNode(X86ISD::SUBUS, dl, Op.getValueType(),
13619 Op.getOperand(1), Op.getOperand(2));
13621 // SSE3/AVX horizontal add/sub intrinsics
13622 case Intrinsic::x86_sse3_hadd_ps:
13623 case Intrinsic::x86_sse3_hadd_pd:
13624 case Intrinsic::x86_avx_hadd_ps_256:
13625 case Intrinsic::x86_avx_hadd_pd_256:
13626 case Intrinsic::x86_sse3_hsub_ps:
13627 case Intrinsic::x86_sse3_hsub_pd:
13628 case Intrinsic::x86_avx_hsub_ps_256:
13629 case Intrinsic::x86_avx_hsub_pd_256:
13630 case Intrinsic::x86_ssse3_phadd_w_128:
13631 case Intrinsic::x86_ssse3_phadd_d_128:
13632 case Intrinsic::x86_avx2_phadd_w:
13633 case Intrinsic::x86_avx2_phadd_d:
13634 case Intrinsic::x86_ssse3_phsub_w_128:
13635 case Intrinsic::x86_ssse3_phsub_d_128:
13636 case Intrinsic::x86_avx2_phsub_w:
13637 case Intrinsic::x86_avx2_phsub_d: {
13640 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
13641 case Intrinsic::x86_sse3_hadd_ps:
13642 case Intrinsic::x86_sse3_hadd_pd:
13643 case Intrinsic::x86_avx_hadd_ps_256:
13644 case Intrinsic::x86_avx_hadd_pd_256:
13645 Opcode = X86ISD::FHADD;
13647 case Intrinsic::x86_sse3_hsub_ps:
13648 case Intrinsic::x86_sse3_hsub_pd:
13649 case Intrinsic::x86_avx_hsub_ps_256:
13650 case Intrinsic::x86_avx_hsub_pd_256:
13651 Opcode = X86ISD::FHSUB;
13653 case Intrinsic::x86_ssse3_phadd_w_128:
13654 case Intrinsic::x86_ssse3_phadd_d_128:
13655 case Intrinsic::x86_avx2_phadd_w:
13656 case Intrinsic::x86_avx2_phadd_d:
13657 Opcode = X86ISD::HADD;
13659 case Intrinsic::x86_ssse3_phsub_w_128:
13660 case Intrinsic::x86_ssse3_phsub_d_128:
13661 case Intrinsic::x86_avx2_phsub_w:
13662 case Intrinsic::x86_avx2_phsub_d:
13663 Opcode = X86ISD::HSUB;
13666 return DAG.getNode(Opcode, dl, Op.getValueType(),
13667 Op.getOperand(1), Op.getOperand(2));
13670 // SSE2/SSE41/AVX2 integer max/min intrinsics.
13671 case Intrinsic::x86_sse2_pmaxu_b:
13672 case Intrinsic::x86_sse41_pmaxuw:
13673 case Intrinsic::x86_sse41_pmaxud:
13674 case Intrinsic::x86_avx2_pmaxu_b:
13675 case Intrinsic::x86_avx2_pmaxu_w:
13676 case Intrinsic::x86_avx2_pmaxu_d:
13677 case Intrinsic::x86_sse2_pminu_b:
13678 case Intrinsic::x86_sse41_pminuw:
13679 case Intrinsic::x86_sse41_pminud:
13680 case Intrinsic::x86_avx2_pminu_b:
13681 case Intrinsic::x86_avx2_pminu_w:
13682 case Intrinsic::x86_avx2_pminu_d:
13683 case Intrinsic::x86_sse41_pmaxsb:
13684 case Intrinsic::x86_sse2_pmaxs_w:
13685 case Intrinsic::x86_sse41_pmaxsd:
13686 case Intrinsic::x86_avx2_pmaxs_b:
13687 case Intrinsic::x86_avx2_pmaxs_w:
13688 case Intrinsic::x86_avx2_pmaxs_d:
13689 case Intrinsic::x86_sse41_pminsb:
13690 case Intrinsic::x86_sse2_pmins_w:
13691 case Intrinsic::x86_sse41_pminsd:
13692 case Intrinsic::x86_avx2_pmins_b:
13693 case Intrinsic::x86_avx2_pmins_w:
13694 case Intrinsic::x86_avx2_pmins_d: {
13697 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
13698 case Intrinsic::x86_sse2_pmaxu_b:
13699 case Intrinsic::x86_sse41_pmaxuw:
13700 case Intrinsic::x86_sse41_pmaxud:
13701 case Intrinsic::x86_avx2_pmaxu_b:
13702 case Intrinsic::x86_avx2_pmaxu_w:
13703 case Intrinsic::x86_avx2_pmaxu_d:
13704 Opcode = X86ISD::UMAX;
13706 case Intrinsic::x86_sse2_pminu_b:
13707 case Intrinsic::x86_sse41_pminuw:
13708 case Intrinsic::x86_sse41_pminud:
13709 case Intrinsic::x86_avx2_pminu_b:
13710 case Intrinsic::x86_avx2_pminu_w:
13711 case Intrinsic::x86_avx2_pminu_d:
13712 Opcode = X86ISD::UMIN;
13714 case Intrinsic::x86_sse41_pmaxsb:
13715 case Intrinsic::x86_sse2_pmaxs_w:
13716 case Intrinsic::x86_sse41_pmaxsd:
13717 case Intrinsic::x86_avx2_pmaxs_b:
13718 case Intrinsic::x86_avx2_pmaxs_w:
13719 case Intrinsic::x86_avx2_pmaxs_d:
13720 Opcode = X86ISD::SMAX;
13722 case Intrinsic::x86_sse41_pminsb:
13723 case Intrinsic::x86_sse2_pmins_w:
13724 case Intrinsic::x86_sse41_pminsd:
13725 case Intrinsic::x86_avx2_pmins_b:
13726 case Intrinsic::x86_avx2_pmins_w:
13727 case Intrinsic::x86_avx2_pmins_d:
13728 Opcode = X86ISD::SMIN;
13731 return DAG.getNode(Opcode, dl, Op.getValueType(),
13732 Op.getOperand(1), Op.getOperand(2));
13735 // SSE/SSE2/AVX floating point max/min intrinsics.
13736 case Intrinsic::x86_sse_max_ps:
13737 case Intrinsic::x86_sse2_max_pd:
13738 case Intrinsic::x86_avx_max_ps_256:
13739 case Intrinsic::x86_avx_max_pd_256:
13740 case Intrinsic::x86_sse_min_ps:
13741 case Intrinsic::x86_sse2_min_pd:
13742 case Intrinsic::x86_avx_min_ps_256:
13743 case Intrinsic::x86_avx_min_pd_256: {
13746 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
13747 case Intrinsic::x86_sse_max_ps:
13748 case Intrinsic::x86_sse2_max_pd:
13749 case Intrinsic::x86_avx_max_ps_256:
13750 case Intrinsic::x86_avx_max_pd_256:
13751 Opcode = X86ISD::FMAX;
13753 case Intrinsic::x86_sse_min_ps:
13754 case Intrinsic::x86_sse2_min_pd:
13755 case Intrinsic::x86_avx_min_ps_256:
13756 case Intrinsic::x86_avx_min_pd_256:
13757 Opcode = X86ISD::FMIN;
13760 return DAG.getNode(Opcode, dl, Op.getValueType(),
13761 Op.getOperand(1), Op.getOperand(2));
13764 // AVX2 variable shift intrinsics
13765 case Intrinsic::x86_avx2_psllv_d:
13766 case Intrinsic::x86_avx2_psllv_q:
13767 case Intrinsic::x86_avx2_psllv_d_256:
13768 case Intrinsic::x86_avx2_psllv_q_256:
13769 case Intrinsic::x86_avx2_psrlv_d:
13770 case Intrinsic::x86_avx2_psrlv_q:
13771 case Intrinsic::x86_avx2_psrlv_d_256:
13772 case Intrinsic::x86_avx2_psrlv_q_256:
13773 case Intrinsic::x86_avx2_psrav_d:
13774 case Intrinsic::x86_avx2_psrav_d_256: {
13777 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
13778 case Intrinsic::x86_avx2_psllv_d:
13779 case Intrinsic::x86_avx2_psllv_q:
13780 case Intrinsic::x86_avx2_psllv_d_256:
13781 case Intrinsic::x86_avx2_psllv_q_256:
13784 case Intrinsic::x86_avx2_psrlv_d:
13785 case Intrinsic::x86_avx2_psrlv_q:
13786 case Intrinsic::x86_avx2_psrlv_d_256:
13787 case Intrinsic::x86_avx2_psrlv_q_256:
13790 case Intrinsic::x86_avx2_psrav_d:
13791 case Intrinsic::x86_avx2_psrav_d_256:
13795 return DAG.getNode(Opcode, dl, Op.getValueType(),
13796 Op.getOperand(1), Op.getOperand(2));
13799 case Intrinsic::x86_sse2_packssdw_128:
13800 case Intrinsic::x86_sse2_packsswb_128:
13801 case Intrinsic::x86_avx2_packssdw:
13802 case Intrinsic::x86_avx2_packsswb:
13803 return DAG.getNode(X86ISD::PACKSS, dl, Op.getValueType(),
13804 Op.getOperand(1), Op.getOperand(2));
13806 case Intrinsic::x86_sse2_packuswb_128:
13807 case Intrinsic::x86_sse41_packusdw:
13808 case Intrinsic::x86_avx2_packuswb:
13809 case Intrinsic::x86_avx2_packusdw:
13810 return DAG.getNode(X86ISD::PACKUS, dl, Op.getValueType(),
13811 Op.getOperand(1), Op.getOperand(2));
13813 case Intrinsic::x86_ssse3_pshuf_b_128:
13814 case Intrinsic::x86_avx2_pshuf_b:
13815 return DAG.getNode(X86ISD::PSHUFB, dl, Op.getValueType(),
13816 Op.getOperand(1), Op.getOperand(2));
13818 case Intrinsic::x86_sse2_pshuf_d:
13819 return DAG.getNode(X86ISD::PSHUFD, dl, Op.getValueType(),
13820 Op.getOperand(1), Op.getOperand(2));
13822 case Intrinsic::x86_sse2_pshufl_w:
13823 return DAG.getNode(X86ISD::PSHUFLW, dl, Op.getValueType(),
13824 Op.getOperand(1), Op.getOperand(2));
13826 case Intrinsic::x86_sse2_pshufh_w:
13827 return DAG.getNode(X86ISD::PSHUFHW, dl, Op.getValueType(),
13828 Op.getOperand(1), Op.getOperand(2));
13830 case Intrinsic::x86_ssse3_psign_b_128:
13831 case Intrinsic::x86_ssse3_psign_w_128:
13832 case Intrinsic::x86_ssse3_psign_d_128:
13833 case Intrinsic::x86_avx2_psign_b:
13834 case Intrinsic::x86_avx2_psign_w:
13835 case Intrinsic::x86_avx2_psign_d:
13836 return DAG.getNode(X86ISD::PSIGN, dl, Op.getValueType(),
13837 Op.getOperand(1), Op.getOperand(2));
13839 case Intrinsic::x86_sse41_insertps:
13840 return DAG.getNode(X86ISD::INSERTPS, dl, Op.getValueType(),
13841 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
13843 case Intrinsic::x86_avx_vperm2f128_ps_256:
13844 case Intrinsic::x86_avx_vperm2f128_pd_256:
13845 case Intrinsic::x86_avx_vperm2f128_si_256:
13846 case Intrinsic::x86_avx2_vperm2i128:
13847 return DAG.getNode(X86ISD::VPERM2X128, dl, Op.getValueType(),
13848 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
13850 case Intrinsic::x86_avx2_permd:
13851 case Intrinsic::x86_avx2_permps:
13852 // Operands intentionally swapped. Mask is last operand to intrinsic,
13853 // but second operand for node/instruction.
13854 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
13855 Op.getOperand(2), Op.getOperand(1));
13857 case Intrinsic::x86_sse_sqrt_ps:
13858 case Intrinsic::x86_sse2_sqrt_pd:
13859 case Intrinsic::x86_avx_sqrt_ps_256:
13860 case Intrinsic::x86_avx_sqrt_pd_256:
13861 return DAG.getNode(ISD::FSQRT, dl, Op.getValueType(), Op.getOperand(1));
13863 // ptest and testp intrinsics. The intrinsic these come from are designed to
13864 // return an integer value, not just an instruction so lower it to the ptest
13865 // or testp pattern and a setcc for the result.
13866 case Intrinsic::x86_sse41_ptestz:
13867 case Intrinsic::x86_sse41_ptestc:
13868 case Intrinsic::x86_sse41_ptestnzc:
13869 case Intrinsic::x86_avx_ptestz_256:
13870 case Intrinsic::x86_avx_ptestc_256:
13871 case Intrinsic::x86_avx_ptestnzc_256:
13872 case Intrinsic::x86_avx_vtestz_ps:
13873 case Intrinsic::x86_avx_vtestc_ps:
13874 case Intrinsic::x86_avx_vtestnzc_ps:
13875 case Intrinsic::x86_avx_vtestz_pd:
13876 case Intrinsic::x86_avx_vtestc_pd:
13877 case Intrinsic::x86_avx_vtestnzc_pd:
13878 case Intrinsic::x86_avx_vtestz_ps_256:
13879 case Intrinsic::x86_avx_vtestc_ps_256:
13880 case Intrinsic::x86_avx_vtestnzc_ps_256:
13881 case Intrinsic::x86_avx_vtestz_pd_256:
13882 case Intrinsic::x86_avx_vtestc_pd_256:
13883 case Intrinsic::x86_avx_vtestnzc_pd_256: {
13884 bool IsTestPacked = false;
13887 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
13888 case Intrinsic::x86_avx_vtestz_ps:
13889 case Intrinsic::x86_avx_vtestz_pd:
13890 case Intrinsic::x86_avx_vtestz_ps_256:
13891 case Intrinsic::x86_avx_vtestz_pd_256:
13892 IsTestPacked = true; // Fallthrough
13893 case Intrinsic::x86_sse41_ptestz:
13894 case Intrinsic::x86_avx_ptestz_256:
13896 X86CC = X86::COND_E;
13898 case Intrinsic::x86_avx_vtestc_ps:
13899 case Intrinsic::x86_avx_vtestc_pd:
13900 case Intrinsic::x86_avx_vtestc_ps_256:
13901 case Intrinsic::x86_avx_vtestc_pd_256:
13902 IsTestPacked = true; // Fallthrough
13903 case Intrinsic::x86_sse41_ptestc:
13904 case Intrinsic::x86_avx_ptestc_256:
13906 X86CC = X86::COND_B;
13908 case Intrinsic::x86_avx_vtestnzc_ps:
13909 case Intrinsic::x86_avx_vtestnzc_pd:
13910 case Intrinsic::x86_avx_vtestnzc_ps_256:
13911 case Intrinsic::x86_avx_vtestnzc_pd_256:
13912 IsTestPacked = true; // Fallthrough
13913 case Intrinsic::x86_sse41_ptestnzc:
13914 case Intrinsic::x86_avx_ptestnzc_256:
13916 X86CC = X86::COND_A;
13920 SDValue LHS = Op.getOperand(1);
13921 SDValue RHS = Op.getOperand(2);
13922 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
13923 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
13924 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
13925 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
13926 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
13928 case Intrinsic::x86_avx512_kortestz_w:
13929 case Intrinsic::x86_avx512_kortestc_w: {
13930 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
13931 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
13932 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
13933 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
13934 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
13935 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
13936 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
13939 // SSE/AVX shift intrinsics
13940 case Intrinsic::x86_sse2_psll_w:
13941 case Intrinsic::x86_sse2_psll_d:
13942 case Intrinsic::x86_sse2_psll_q:
13943 case Intrinsic::x86_avx2_psll_w:
13944 case Intrinsic::x86_avx2_psll_d:
13945 case Intrinsic::x86_avx2_psll_q:
13946 case Intrinsic::x86_sse2_psrl_w:
13947 case Intrinsic::x86_sse2_psrl_d:
13948 case Intrinsic::x86_sse2_psrl_q:
13949 case Intrinsic::x86_avx2_psrl_w:
13950 case Intrinsic::x86_avx2_psrl_d:
13951 case Intrinsic::x86_avx2_psrl_q:
13952 case Intrinsic::x86_sse2_psra_w:
13953 case Intrinsic::x86_sse2_psra_d:
13954 case Intrinsic::x86_avx2_psra_w:
13955 case Intrinsic::x86_avx2_psra_d: {
13958 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
13959 case Intrinsic::x86_sse2_psll_w:
13960 case Intrinsic::x86_sse2_psll_d:
13961 case Intrinsic::x86_sse2_psll_q:
13962 case Intrinsic::x86_avx2_psll_w:
13963 case Intrinsic::x86_avx2_psll_d:
13964 case Intrinsic::x86_avx2_psll_q:
13965 Opcode = X86ISD::VSHL;
13967 case Intrinsic::x86_sse2_psrl_w:
13968 case Intrinsic::x86_sse2_psrl_d:
13969 case Intrinsic::x86_sse2_psrl_q:
13970 case Intrinsic::x86_avx2_psrl_w:
13971 case Intrinsic::x86_avx2_psrl_d:
13972 case Intrinsic::x86_avx2_psrl_q:
13973 Opcode = X86ISD::VSRL;
13975 case Intrinsic::x86_sse2_psra_w:
13976 case Intrinsic::x86_sse2_psra_d:
13977 case Intrinsic::x86_avx2_psra_w:
13978 case Intrinsic::x86_avx2_psra_d:
13979 Opcode = X86ISD::VSRA;
13982 return DAG.getNode(Opcode, dl, Op.getValueType(),
13983 Op.getOperand(1), Op.getOperand(2));
13986 // SSE/AVX immediate shift intrinsics
13987 case Intrinsic::x86_sse2_pslli_w:
13988 case Intrinsic::x86_sse2_pslli_d:
13989 case Intrinsic::x86_sse2_pslli_q:
13990 case Intrinsic::x86_avx2_pslli_w:
13991 case Intrinsic::x86_avx2_pslli_d:
13992 case Intrinsic::x86_avx2_pslli_q:
13993 case Intrinsic::x86_sse2_psrli_w:
13994 case Intrinsic::x86_sse2_psrli_d:
13995 case Intrinsic::x86_sse2_psrli_q:
13996 case Intrinsic::x86_avx2_psrli_w:
13997 case Intrinsic::x86_avx2_psrli_d:
13998 case Intrinsic::x86_avx2_psrli_q:
13999 case Intrinsic::x86_sse2_psrai_w:
14000 case Intrinsic::x86_sse2_psrai_d:
14001 case Intrinsic::x86_avx2_psrai_w:
14002 case Intrinsic::x86_avx2_psrai_d: {
14005 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
14006 case Intrinsic::x86_sse2_pslli_w:
14007 case Intrinsic::x86_sse2_pslli_d:
14008 case Intrinsic::x86_sse2_pslli_q:
14009 case Intrinsic::x86_avx2_pslli_w:
14010 case Intrinsic::x86_avx2_pslli_d:
14011 case Intrinsic::x86_avx2_pslli_q:
14012 Opcode = X86ISD::VSHLI;
14014 case Intrinsic::x86_sse2_psrli_w:
14015 case Intrinsic::x86_sse2_psrli_d:
14016 case Intrinsic::x86_sse2_psrli_q:
14017 case Intrinsic::x86_avx2_psrli_w:
14018 case Intrinsic::x86_avx2_psrli_d:
14019 case Intrinsic::x86_avx2_psrli_q:
14020 Opcode = X86ISD::VSRLI;
14022 case Intrinsic::x86_sse2_psrai_w:
14023 case Intrinsic::x86_sse2_psrai_d:
14024 case Intrinsic::x86_avx2_psrai_w:
14025 case Intrinsic::x86_avx2_psrai_d:
14026 Opcode = X86ISD::VSRAI;
14029 return getTargetVShiftNode(Opcode, dl, Op.getSimpleValueType(),
14030 Op.getOperand(1), Op.getOperand(2), DAG);
14033 case Intrinsic::x86_sse42_pcmpistria128:
14034 case Intrinsic::x86_sse42_pcmpestria128:
14035 case Intrinsic::x86_sse42_pcmpistric128:
14036 case Intrinsic::x86_sse42_pcmpestric128:
14037 case Intrinsic::x86_sse42_pcmpistrio128:
14038 case Intrinsic::x86_sse42_pcmpestrio128:
14039 case Intrinsic::x86_sse42_pcmpistris128:
14040 case Intrinsic::x86_sse42_pcmpestris128:
14041 case Intrinsic::x86_sse42_pcmpistriz128:
14042 case Intrinsic::x86_sse42_pcmpestriz128: {
14046 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
14047 case Intrinsic::x86_sse42_pcmpistria128:
14048 Opcode = X86ISD::PCMPISTRI;
14049 X86CC = X86::COND_A;
14051 case Intrinsic::x86_sse42_pcmpestria128:
14052 Opcode = X86ISD::PCMPESTRI;
14053 X86CC = X86::COND_A;
14055 case Intrinsic::x86_sse42_pcmpistric128:
14056 Opcode = X86ISD::PCMPISTRI;
14057 X86CC = X86::COND_B;
14059 case Intrinsic::x86_sse42_pcmpestric128:
14060 Opcode = X86ISD::PCMPESTRI;
14061 X86CC = X86::COND_B;
14063 case Intrinsic::x86_sse42_pcmpistrio128:
14064 Opcode = X86ISD::PCMPISTRI;
14065 X86CC = X86::COND_O;
14067 case Intrinsic::x86_sse42_pcmpestrio128:
14068 Opcode = X86ISD::PCMPESTRI;
14069 X86CC = X86::COND_O;
14071 case Intrinsic::x86_sse42_pcmpistris128:
14072 Opcode = X86ISD::PCMPISTRI;
14073 X86CC = X86::COND_S;
14075 case Intrinsic::x86_sse42_pcmpestris128:
14076 Opcode = X86ISD::PCMPESTRI;
14077 X86CC = X86::COND_S;
14079 case Intrinsic::x86_sse42_pcmpistriz128:
14080 Opcode = X86ISD::PCMPISTRI;
14081 X86CC = X86::COND_E;
14083 case Intrinsic::x86_sse42_pcmpestriz128:
14084 Opcode = X86ISD::PCMPESTRI;
14085 X86CC = X86::COND_E;
14088 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
14089 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14090 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
14091 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14092 DAG.getConstant(X86CC, MVT::i8),
14093 SDValue(PCMP.getNode(), 1));
14094 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14097 case Intrinsic::x86_sse42_pcmpistri128:
14098 case Intrinsic::x86_sse42_pcmpestri128: {
14100 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
14101 Opcode = X86ISD::PCMPISTRI;
14103 Opcode = X86ISD::PCMPESTRI;
14105 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
14106 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14107 return DAG.getNode(Opcode, dl, VTs, NewOps);
14109 case Intrinsic::x86_fma_vfmadd_ps:
14110 case Intrinsic::x86_fma_vfmadd_pd:
14111 case Intrinsic::x86_fma_vfmsub_ps:
14112 case Intrinsic::x86_fma_vfmsub_pd:
14113 case Intrinsic::x86_fma_vfnmadd_ps:
14114 case Intrinsic::x86_fma_vfnmadd_pd:
14115 case Intrinsic::x86_fma_vfnmsub_ps:
14116 case Intrinsic::x86_fma_vfnmsub_pd:
14117 case Intrinsic::x86_fma_vfmaddsub_ps:
14118 case Intrinsic::x86_fma_vfmaddsub_pd:
14119 case Intrinsic::x86_fma_vfmsubadd_ps:
14120 case Intrinsic::x86_fma_vfmsubadd_pd:
14121 case Intrinsic::x86_fma_vfmadd_ps_256:
14122 case Intrinsic::x86_fma_vfmadd_pd_256:
14123 case Intrinsic::x86_fma_vfmsub_ps_256:
14124 case Intrinsic::x86_fma_vfmsub_pd_256:
14125 case Intrinsic::x86_fma_vfnmadd_ps_256:
14126 case Intrinsic::x86_fma_vfnmadd_pd_256:
14127 case Intrinsic::x86_fma_vfnmsub_ps_256:
14128 case Intrinsic::x86_fma_vfnmsub_pd_256:
14129 case Intrinsic::x86_fma_vfmaddsub_ps_256:
14130 case Intrinsic::x86_fma_vfmaddsub_pd_256:
14131 case Intrinsic::x86_fma_vfmsubadd_ps_256:
14132 case Intrinsic::x86_fma_vfmsubadd_pd_256:
14133 case Intrinsic::x86_fma_vfmadd_ps_512:
14134 case Intrinsic::x86_fma_vfmadd_pd_512:
14135 case Intrinsic::x86_fma_vfmsub_ps_512:
14136 case Intrinsic::x86_fma_vfmsub_pd_512:
14137 case Intrinsic::x86_fma_vfnmadd_ps_512:
14138 case Intrinsic::x86_fma_vfnmadd_pd_512:
14139 case Intrinsic::x86_fma_vfnmsub_ps_512:
14140 case Intrinsic::x86_fma_vfnmsub_pd_512:
14141 case Intrinsic::x86_fma_vfmaddsub_ps_512:
14142 case Intrinsic::x86_fma_vfmaddsub_pd_512:
14143 case Intrinsic::x86_fma_vfmsubadd_ps_512:
14144 case Intrinsic::x86_fma_vfmsubadd_pd_512: {
14147 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
14148 case Intrinsic::x86_fma_vfmadd_ps:
14149 case Intrinsic::x86_fma_vfmadd_pd:
14150 case Intrinsic::x86_fma_vfmadd_ps_256:
14151 case Intrinsic::x86_fma_vfmadd_pd_256:
14152 case Intrinsic::x86_fma_vfmadd_ps_512:
14153 case Intrinsic::x86_fma_vfmadd_pd_512:
14154 Opc = X86ISD::FMADD;
14156 case Intrinsic::x86_fma_vfmsub_ps:
14157 case Intrinsic::x86_fma_vfmsub_pd:
14158 case Intrinsic::x86_fma_vfmsub_ps_256:
14159 case Intrinsic::x86_fma_vfmsub_pd_256:
14160 case Intrinsic::x86_fma_vfmsub_ps_512:
14161 case Intrinsic::x86_fma_vfmsub_pd_512:
14162 Opc = X86ISD::FMSUB;
14164 case Intrinsic::x86_fma_vfnmadd_ps:
14165 case Intrinsic::x86_fma_vfnmadd_pd:
14166 case Intrinsic::x86_fma_vfnmadd_ps_256:
14167 case Intrinsic::x86_fma_vfnmadd_pd_256:
14168 case Intrinsic::x86_fma_vfnmadd_ps_512:
14169 case Intrinsic::x86_fma_vfnmadd_pd_512:
14170 Opc = X86ISD::FNMADD;
14172 case Intrinsic::x86_fma_vfnmsub_ps:
14173 case Intrinsic::x86_fma_vfnmsub_pd:
14174 case Intrinsic::x86_fma_vfnmsub_ps_256:
14175 case Intrinsic::x86_fma_vfnmsub_pd_256:
14176 case Intrinsic::x86_fma_vfnmsub_ps_512:
14177 case Intrinsic::x86_fma_vfnmsub_pd_512:
14178 Opc = X86ISD::FNMSUB;
14180 case Intrinsic::x86_fma_vfmaddsub_ps:
14181 case Intrinsic::x86_fma_vfmaddsub_pd:
14182 case Intrinsic::x86_fma_vfmaddsub_ps_256:
14183 case Intrinsic::x86_fma_vfmaddsub_pd_256:
14184 case Intrinsic::x86_fma_vfmaddsub_ps_512:
14185 case Intrinsic::x86_fma_vfmaddsub_pd_512:
14186 Opc = X86ISD::FMADDSUB;
14188 case Intrinsic::x86_fma_vfmsubadd_ps:
14189 case Intrinsic::x86_fma_vfmsubadd_pd:
14190 case Intrinsic::x86_fma_vfmsubadd_ps_256:
14191 case Intrinsic::x86_fma_vfmsubadd_pd_256:
14192 case Intrinsic::x86_fma_vfmsubadd_ps_512:
14193 case Intrinsic::x86_fma_vfmsubadd_pd_512:
14194 Opc = X86ISD::FMSUBADD;
14198 return DAG.getNode(Opc, dl, Op.getValueType(), Op.getOperand(1),
14199 Op.getOperand(2), Op.getOperand(3));
14204 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14205 SDValue Src, SDValue Mask, SDValue Base,
14206 SDValue Index, SDValue ScaleOp, SDValue Chain,
14207 const X86Subtarget * Subtarget) {
14209 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14210 assert(C && "Invalid scale type");
14211 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14212 EVT MaskVT = MVT::getVectorVT(MVT::i1,
14213 Index.getSimpleValueType().getVectorNumElements());
14215 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14217 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14219 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14220 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
14221 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14222 SDValue Segment = DAG.getRegister(0, MVT::i32);
14223 if (Src.getOpcode() == ISD::UNDEF)
14224 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
14225 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
14226 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
14227 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
14228 return DAG.getMergeValues(RetOps, dl);
14231 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14232 SDValue Src, SDValue Mask, SDValue Base,
14233 SDValue Index, SDValue ScaleOp, SDValue Chain) {
14235 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14236 assert(C && "Invalid scale type");
14237 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14238 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14239 SDValue Segment = DAG.getRegister(0, MVT::i32);
14240 EVT MaskVT = MVT::getVectorVT(MVT::i1,
14241 Index.getSimpleValueType().getVectorNumElements());
14243 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14245 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14247 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14248 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
14249 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
14250 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
14251 return SDValue(Res, 1);
14254 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14255 SDValue Mask, SDValue Base, SDValue Index,
14256 SDValue ScaleOp, SDValue Chain) {
14258 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14259 assert(C && "Invalid scale type");
14260 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14261 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14262 SDValue Segment = DAG.getRegister(0, MVT::i32);
14264 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
14266 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14268 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14270 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14271 //SDVTList VTs = DAG.getVTList(MVT::Other);
14272 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
14273 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
14274 return SDValue(Res, 0);
14277 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
14278 // read performance monitor counters (x86_rdpmc).
14279 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
14280 SelectionDAG &DAG, const X86Subtarget *Subtarget,
14281 SmallVectorImpl<SDValue> &Results) {
14282 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
14283 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
14286 // The ECX register is used to select the index of the performance counter
14288 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
14290 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
14292 // Reads the content of a 64-bit performance counter and returns it in the
14293 // registers EDX:EAX.
14294 if (Subtarget->is64Bit()) {
14295 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
14296 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
14299 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
14300 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
14303 Chain = HI.getValue(1);
14305 if (Subtarget->is64Bit()) {
14306 // The EAX register is loaded with the low-order 32 bits. The EDX register
14307 // is loaded with the supported high-order bits of the counter.
14308 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
14309 DAG.getConstant(32, MVT::i8));
14310 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
14311 Results.push_back(Chain);
14315 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
14316 SDValue Ops[] = { LO, HI };
14317 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
14318 Results.push_back(Pair);
14319 Results.push_back(Chain);
14322 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
14323 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
14324 // also used to custom lower READCYCLECOUNTER nodes.
14325 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
14326 SelectionDAG &DAG, const X86Subtarget *Subtarget,
14327 SmallVectorImpl<SDValue> &Results) {
14328 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
14329 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
14332 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
14333 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
14334 // and the EAX register is loaded with the low-order 32 bits.
14335 if (Subtarget->is64Bit()) {
14336 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
14337 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
14340 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
14341 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
14344 SDValue Chain = HI.getValue(1);
14346 if (Opcode == X86ISD::RDTSCP_DAG) {
14347 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
14349 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
14350 // the ECX register. Add 'ecx' explicitly to the chain.
14351 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
14353 // Explicitly store the content of ECX at the location passed in input
14354 // to the 'rdtscp' intrinsic.
14355 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
14356 MachinePointerInfo(), false, false, 0);
14359 if (Subtarget->is64Bit()) {
14360 // The EDX register is loaded with the high-order 32 bits of the MSR, and
14361 // the EAX register is loaded with the low-order 32 bits.
14362 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
14363 DAG.getConstant(32, MVT::i8));
14364 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
14365 Results.push_back(Chain);
14369 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
14370 SDValue Ops[] = { LO, HI };
14371 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
14372 Results.push_back(Pair);
14373 Results.push_back(Chain);
14376 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
14377 SelectionDAG &DAG) {
14378 SmallVector<SDValue, 2> Results;
14380 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
14382 return DAG.getMergeValues(Results, DL);
14385 enum IntrinsicType {
14386 GATHER, SCATTER, PREFETCH, RDSEED, RDRAND, RDPMC, RDTSC, XTEST
14389 struct IntrinsicData {
14390 IntrinsicData(IntrinsicType IType, unsigned IOpc0, unsigned IOpc1)
14391 :Type(IType), Opc0(IOpc0), Opc1(IOpc1) {}
14392 IntrinsicType Type;
14397 std::map < unsigned, IntrinsicData> IntrMap;
14398 static void InitIntinsicsMap() {
14399 static bool Initialized = false;
14402 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qps_512,
14403 IntrinsicData(GATHER, X86::VGATHERQPSZrm, 0)));
14404 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qps_512,
14405 IntrinsicData(GATHER, X86::VGATHERQPSZrm, 0)));
14406 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qpd_512,
14407 IntrinsicData(GATHER, X86::VGATHERQPDZrm, 0)));
14408 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_dpd_512,
14409 IntrinsicData(GATHER, X86::VGATHERDPDZrm, 0)));
14410 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_dps_512,
14411 IntrinsicData(GATHER, X86::VGATHERDPSZrm, 0)));
14412 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qpi_512,
14413 IntrinsicData(GATHER, X86::VPGATHERQDZrm, 0)));
14414 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qpq_512,
14415 IntrinsicData(GATHER, X86::VPGATHERQQZrm, 0)));
14416 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_dpi_512,
14417 IntrinsicData(GATHER, X86::VPGATHERDDZrm, 0)));
14418 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_dpq_512,
14419 IntrinsicData(GATHER, X86::VPGATHERDQZrm, 0)));
14421 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_qps_512,
14422 IntrinsicData(SCATTER, X86::VSCATTERQPSZmr, 0)));
14423 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_qpd_512,
14424 IntrinsicData(SCATTER, X86::VSCATTERQPDZmr, 0)));
14425 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_dpd_512,
14426 IntrinsicData(SCATTER, X86::VSCATTERDPDZmr, 0)));
14427 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_dps_512,
14428 IntrinsicData(SCATTER, X86::VSCATTERDPSZmr, 0)));
14429 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_qpi_512,
14430 IntrinsicData(SCATTER, X86::VPSCATTERQDZmr, 0)));
14431 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_qpq_512,
14432 IntrinsicData(SCATTER, X86::VPSCATTERQQZmr, 0)));
14433 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_dpi_512,
14434 IntrinsicData(SCATTER, X86::VPSCATTERDDZmr, 0)));
14435 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_dpq_512,
14436 IntrinsicData(SCATTER, X86::VPSCATTERDQZmr, 0)));
14438 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gatherpf_qps_512,
14439 IntrinsicData(PREFETCH, X86::VGATHERPF0QPSm,
14440 X86::VGATHERPF1QPSm)));
14441 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gatherpf_qpd_512,
14442 IntrinsicData(PREFETCH, X86::VGATHERPF0QPDm,
14443 X86::VGATHERPF1QPDm)));
14444 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gatherpf_dpd_512,
14445 IntrinsicData(PREFETCH, X86::VGATHERPF0DPDm,
14446 X86::VGATHERPF1DPDm)));
14447 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gatherpf_dps_512,
14448 IntrinsicData(PREFETCH, X86::VGATHERPF0DPSm,
14449 X86::VGATHERPF1DPSm)));
14450 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatterpf_qps_512,
14451 IntrinsicData(PREFETCH, X86::VSCATTERPF0QPSm,
14452 X86::VSCATTERPF1QPSm)));
14453 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatterpf_qpd_512,
14454 IntrinsicData(PREFETCH, X86::VSCATTERPF0QPDm,
14455 X86::VSCATTERPF1QPDm)));
14456 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatterpf_dpd_512,
14457 IntrinsicData(PREFETCH, X86::VSCATTERPF0DPDm,
14458 X86::VSCATTERPF1DPDm)));
14459 IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatterpf_dps_512,
14460 IntrinsicData(PREFETCH, X86::VSCATTERPF0DPSm,
14461 X86::VSCATTERPF1DPSm)));
14462 IntrMap.insert(std::make_pair(Intrinsic::x86_rdrand_16,
14463 IntrinsicData(RDRAND, X86ISD::RDRAND, 0)));
14464 IntrMap.insert(std::make_pair(Intrinsic::x86_rdrand_32,
14465 IntrinsicData(RDRAND, X86ISD::RDRAND, 0)));
14466 IntrMap.insert(std::make_pair(Intrinsic::x86_rdrand_64,
14467 IntrinsicData(RDRAND, X86ISD::RDRAND, 0)));
14468 IntrMap.insert(std::make_pair(Intrinsic::x86_rdseed_16,
14469 IntrinsicData(RDSEED, X86ISD::RDSEED, 0)));
14470 IntrMap.insert(std::make_pair(Intrinsic::x86_rdseed_32,
14471 IntrinsicData(RDSEED, X86ISD::RDSEED, 0)));
14472 IntrMap.insert(std::make_pair(Intrinsic::x86_rdseed_64,
14473 IntrinsicData(RDSEED, X86ISD::RDSEED, 0)));
14474 IntrMap.insert(std::make_pair(Intrinsic::x86_xtest,
14475 IntrinsicData(XTEST, X86ISD::XTEST, 0)));
14476 IntrMap.insert(std::make_pair(Intrinsic::x86_rdtsc,
14477 IntrinsicData(RDTSC, X86ISD::RDTSC_DAG, 0)));
14478 IntrMap.insert(std::make_pair(Intrinsic::x86_rdtscp,
14479 IntrinsicData(RDTSC, X86ISD::RDTSCP_DAG, 0)));
14480 IntrMap.insert(std::make_pair(Intrinsic::x86_rdpmc,
14481 IntrinsicData(RDPMC, X86ISD::RDPMC_DAG, 0)));
14482 Initialized = true;
14485 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14486 SelectionDAG &DAG) {
14487 InitIntinsicsMap();
14488 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
14489 std::map < unsigned, IntrinsicData>::const_iterator itr = IntrMap.find(IntNo);
14490 if (itr == IntrMap.end())
14494 IntrinsicData Intr = itr->second;
14495 switch(Intr.Type) {
14498 // Emit the node with the right value type.
14499 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
14500 SDValue Result = DAG.getNode(Intr.Opc0, dl, VTs, Op.getOperand(0));
14502 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
14503 // Otherwise return the value from Rand, which is always 0, casted to i32.
14504 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
14505 DAG.getConstant(1, Op->getValueType(1)),
14506 DAG.getConstant(X86::COND_B, MVT::i32),
14507 SDValue(Result.getNode(), 1) };
14508 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
14509 DAG.getVTList(Op->getValueType(1), MVT::Glue),
14512 // Return { result, isValid, chain }.
14513 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
14514 SDValue(Result.getNode(), 2));
14517 //gather(v1, mask, index, base, scale);
14518 SDValue Chain = Op.getOperand(0);
14519 SDValue Src = Op.getOperand(2);
14520 SDValue Base = Op.getOperand(3);
14521 SDValue Index = Op.getOperand(4);
14522 SDValue Mask = Op.getOperand(5);
14523 SDValue Scale = Op.getOperand(6);
14524 return getGatherNode(Intr.Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
14528 //scatter(base, mask, index, v1, scale);
14529 SDValue Chain = Op.getOperand(0);
14530 SDValue Base = Op.getOperand(2);
14531 SDValue Mask = Op.getOperand(3);
14532 SDValue Index = Op.getOperand(4);
14533 SDValue Src = Op.getOperand(5);
14534 SDValue Scale = Op.getOperand(6);
14535 return getScatterNode(Intr.Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
14538 SDValue Hint = Op.getOperand(6);
14540 if (dyn_cast<ConstantSDNode> (Hint) == nullptr ||
14541 (HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1)
14542 llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1");
14543 unsigned Opcode = (HintVal ? Intr.Opc1 : Intr.Opc0);
14544 SDValue Chain = Op.getOperand(0);
14545 SDValue Mask = Op.getOperand(2);
14546 SDValue Index = Op.getOperand(3);
14547 SDValue Base = Op.getOperand(4);
14548 SDValue Scale = Op.getOperand(5);
14549 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
14551 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
14553 SmallVector<SDValue, 2> Results;
14554 getReadTimeStampCounter(Op.getNode(), dl, Intr.Opc0, DAG, Subtarget, Results);
14555 return DAG.getMergeValues(Results, dl);
14557 // Read Performance Monitoring Counters.
14559 SmallVector<SDValue, 2> Results;
14560 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
14561 return DAG.getMergeValues(Results, dl);
14563 // XTEST intrinsics.
14565 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
14566 SDValue InTrans = DAG.getNode(X86ISD::XTEST, dl, VTs, Op.getOperand(0));
14567 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14568 DAG.getConstant(X86::COND_NE, MVT::i8),
14570 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
14571 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
14572 Ret, SDValue(InTrans.getNode(), 1));
14575 llvm_unreachable("Unknown Intrinsic Type");
14578 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
14579 SelectionDAG &DAG) const {
14580 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
14581 MFI->setReturnAddressIsTaken(true);
14583 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
14586 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14588 EVT PtrVT = getPointerTy();
14591 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
14592 const X86RegisterInfo *RegInfo =
14593 static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
14594 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
14595 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
14596 DAG.getNode(ISD::ADD, dl, PtrVT,
14597 FrameAddr, Offset),
14598 MachinePointerInfo(), false, false, false, 0);
14601 // Just load the return address.
14602 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
14603 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
14604 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
14607 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
14608 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
14609 MFI->setFrameAddressIsTaken(true);
14611 EVT VT = Op.getValueType();
14612 SDLoc dl(Op); // FIXME probably not meaningful
14613 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14614 const X86RegisterInfo *RegInfo =
14615 static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
14616 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
14617 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
14618 (FrameReg == X86::EBP && VT == MVT::i32)) &&
14619 "Invalid Frame Register!");
14620 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
14622 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
14623 MachinePointerInfo(),
14624 false, false, false, 0);
14628 // FIXME? Maybe this could be a TableGen attribute on some registers and
14629 // this table could be generated automatically from RegInfo.
14630 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
14632 unsigned Reg = StringSwitch<unsigned>(RegName)
14633 .Case("esp", X86::ESP)
14634 .Case("rsp", X86::RSP)
14638 report_fatal_error("Invalid register name global variable");
14641 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
14642 SelectionDAG &DAG) const {
14643 const X86RegisterInfo *RegInfo =
14644 static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
14645 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
14648 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
14649 SDValue Chain = Op.getOperand(0);
14650 SDValue Offset = Op.getOperand(1);
14651 SDValue Handler = Op.getOperand(2);
14654 EVT PtrVT = getPointerTy();
14655 const X86RegisterInfo *RegInfo =
14656 static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
14657 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
14658 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
14659 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
14660 "Invalid Frame Register!");
14661 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
14662 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
14664 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
14665 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
14666 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
14667 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
14669 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
14671 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
14672 DAG.getRegister(StoreAddrReg, PtrVT));
14675 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
14676 SelectionDAG &DAG) const {
14678 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
14679 DAG.getVTList(MVT::i32, MVT::Other),
14680 Op.getOperand(0), Op.getOperand(1));
14683 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
14684 SelectionDAG &DAG) const {
14686 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
14687 Op.getOperand(0), Op.getOperand(1));
14690 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
14691 return Op.getOperand(0);
14694 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
14695 SelectionDAG &DAG) const {
14696 SDValue Root = Op.getOperand(0);
14697 SDValue Trmp = Op.getOperand(1); // trampoline
14698 SDValue FPtr = Op.getOperand(2); // nested function
14699 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
14702 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
14703 const TargetRegisterInfo* TRI = DAG.getTarget().getRegisterInfo();
14705 if (Subtarget->is64Bit()) {
14706 SDValue OutChains[6];
14708 // Large code-model.
14709 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
14710 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
14712 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
14713 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
14715 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
14717 // Load the pointer to the nested function into R11.
14718 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
14719 SDValue Addr = Trmp;
14720 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
14721 Addr, MachinePointerInfo(TrmpAddr),
14724 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
14725 DAG.getConstant(2, MVT::i64));
14726 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
14727 MachinePointerInfo(TrmpAddr, 2),
14730 // Load the 'nest' parameter value into R10.
14731 // R10 is specified in X86CallingConv.td
14732 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
14733 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
14734 DAG.getConstant(10, MVT::i64));
14735 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
14736 Addr, MachinePointerInfo(TrmpAddr, 10),
14739 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
14740 DAG.getConstant(12, MVT::i64));
14741 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
14742 MachinePointerInfo(TrmpAddr, 12),
14745 // Jump to the nested function.
14746 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
14747 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
14748 DAG.getConstant(20, MVT::i64));
14749 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
14750 Addr, MachinePointerInfo(TrmpAddr, 20),
14753 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
14754 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
14755 DAG.getConstant(22, MVT::i64));
14756 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
14757 MachinePointerInfo(TrmpAddr, 22),
14760 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
14762 const Function *Func =
14763 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
14764 CallingConv::ID CC = Func->getCallingConv();
14769 llvm_unreachable("Unsupported calling convention");
14770 case CallingConv::C:
14771 case CallingConv::X86_StdCall: {
14772 // Pass 'nest' parameter in ECX.
14773 // Must be kept in sync with X86CallingConv.td
14774 NestReg = X86::ECX;
14776 // Check that ECX wasn't needed by an 'inreg' parameter.
14777 FunctionType *FTy = Func->getFunctionType();
14778 const AttributeSet &Attrs = Func->getAttributes();
14780 if (!Attrs.isEmpty() && !Func->isVarArg()) {
14781 unsigned InRegCount = 0;
14784 for (FunctionType::param_iterator I = FTy->param_begin(),
14785 E = FTy->param_end(); I != E; ++I, ++Idx)
14786 if (Attrs.hasAttribute(Idx, Attribute::InReg))
14787 // FIXME: should only count parameters that are lowered to integers.
14788 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
14790 if (InRegCount > 2) {
14791 report_fatal_error("Nest register in use - reduce number of inreg"
14797 case CallingConv::X86_FastCall:
14798 case CallingConv::X86_ThisCall:
14799 case CallingConv::Fast:
14800 // Pass 'nest' parameter in EAX.
14801 // Must be kept in sync with X86CallingConv.td
14802 NestReg = X86::EAX;
14806 SDValue OutChains[4];
14807 SDValue Addr, Disp;
14809 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
14810 DAG.getConstant(10, MVT::i32));
14811 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
14813 // This is storing the opcode for MOV32ri.
14814 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
14815 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
14816 OutChains[0] = DAG.getStore(Root, dl,
14817 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
14818 Trmp, MachinePointerInfo(TrmpAddr),
14821 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
14822 DAG.getConstant(1, MVT::i32));
14823 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
14824 MachinePointerInfo(TrmpAddr, 1),
14827 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
14828 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
14829 DAG.getConstant(5, MVT::i32));
14830 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
14831 MachinePointerInfo(TrmpAddr, 5),
14834 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
14835 DAG.getConstant(6, MVT::i32));
14836 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
14837 MachinePointerInfo(TrmpAddr, 6),
14840 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
14844 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
14845 SelectionDAG &DAG) const {
14847 The rounding mode is in bits 11:10 of FPSR, and has the following
14849 00 Round to nearest
14854 FLT_ROUNDS, on the other hand, expects the following:
14861 To perform the conversion, we do:
14862 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
14865 MachineFunction &MF = DAG.getMachineFunction();
14866 const TargetMachine &TM = MF.getTarget();
14867 const TargetFrameLowering &TFI = *TM.getFrameLowering();
14868 unsigned StackAlignment = TFI.getStackAlignment();
14869 MVT VT = Op.getSimpleValueType();
14872 // Save FP Control Word to stack slot
14873 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
14874 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14876 MachineMemOperand *MMO =
14877 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14878 MachineMemOperand::MOStore, 2, 2);
14880 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
14881 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
14882 DAG.getVTList(MVT::Other),
14883 Ops, MVT::i16, MMO);
14885 // Load FP Control Word from stack slot
14886 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
14887 MachinePointerInfo(), false, false, false, 0);
14889 // Transform as necessary
14891 DAG.getNode(ISD::SRL, DL, MVT::i16,
14892 DAG.getNode(ISD::AND, DL, MVT::i16,
14893 CWD, DAG.getConstant(0x800, MVT::i16)),
14894 DAG.getConstant(11, MVT::i8));
14896 DAG.getNode(ISD::SRL, DL, MVT::i16,
14897 DAG.getNode(ISD::AND, DL, MVT::i16,
14898 CWD, DAG.getConstant(0x400, MVT::i16)),
14899 DAG.getConstant(9, MVT::i8));
14902 DAG.getNode(ISD::AND, DL, MVT::i16,
14903 DAG.getNode(ISD::ADD, DL, MVT::i16,
14904 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
14905 DAG.getConstant(1, MVT::i16)),
14906 DAG.getConstant(3, MVT::i16));
14908 return DAG.getNode((VT.getSizeInBits() < 16 ?
14909 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
14912 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
14913 MVT VT = Op.getSimpleValueType();
14915 unsigned NumBits = VT.getSizeInBits();
14918 Op = Op.getOperand(0);
14919 if (VT == MVT::i8) {
14920 // Zero extend to i32 since there is not an i8 bsr.
14922 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
14925 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
14926 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
14927 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
14929 // If src is zero (i.e. bsr sets ZF), returns NumBits.
14932 DAG.getConstant(NumBits+NumBits-1, OpVT),
14933 DAG.getConstant(X86::COND_E, MVT::i8),
14936 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
14938 // Finally xor with NumBits-1.
14939 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
14942 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
14946 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
14947 MVT VT = Op.getSimpleValueType();
14949 unsigned NumBits = VT.getSizeInBits();
14952 Op = Op.getOperand(0);
14953 if (VT == MVT::i8) {
14954 // Zero extend to i32 since there is not an i8 bsr.
14956 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
14959 // Issue a bsr (scan bits in reverse).
14960 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
14961 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
14963 // And xor with NumBits-1.
14964 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
14967 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
14971 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
14972 MVT VT = Op.getSimpleValueType();
14973 unsigned NumBits = VT.getSizeInBits();
14975 Op = Op.getOperand(0);
14977 // Issue a bsf (scan bits forward) which also sets EFLAGS.
14978 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
14979 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
14981 // If src is zero (i.e. bsf sets ZF), returns NumBits.
14984 DAG.getConstant(NumBits, VT),
14985 DAG.getConstant(X86::COND_E, MVT::i8),
14988 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
14991 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
14992 // ones, and then concatenate the result back.
14993 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
14994 MVT VT = Op.getSimpleValueType();
14996 assert(VT.is256BitVector() && VT.isInteger() &&
14997 "Unsupported value type for operation");
14999 unsigned NumElems = VT.getVectorNumElements();
15002 // Extract the LHS vectors
15003 SDValue LHS = Op.getOperand(0);
15004 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15005 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15007 // Extract the RHS vectors
15008 SDValue RHS = Op.getOperand(1);
15009 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
15010 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
15012 MVT EltVT = VT.getVectorElementType();
15013 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15015 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15016 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
15017 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
15020 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
15021 assert(Op.getSimpleValueType().is256BitVector() &&
15022 Op.getSimpleValueType().isInteger() &&
15023 "Only handle AVX 256-bit vector integer operation");
15024 return Lower256IntArith(Op, DAG);
15027 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
15028 assert(Op.getSimpleValueType().is256BitVector() &&
15029 Op.getSimpleValueType().isInteger() &&
15030 "Only handle AVX 256-bit vector integer operation");
15031 return Lower256IntArith(Op, DAG);
15034 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
15035 SelectionDAG &DAG) {
15037 MVT VT = Op.getSimpleValueType();
15039 // Decompose 256-bit ops into smaller 128-bit ops.
15040 if (VT.is256BitVector() && !Subtarget->hasInt256())
15041 return Lower256IntArith(Op, DAG);
15043 SDValue A = Op.getOperand(0);
15044 SDValue B = Op.getOperand(1);
15046 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
15047 if (VT == MVT::v4i32) {
15048 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
15049 "Should not custom lower when pmuldq is available!");
15051 // Extract the odd parts.
15052 static const int UnpackMask[] = { 1, -1, 3, -1 };
15053 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
15054 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
15056 // Multiply the even parts.
15057 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
15058 // Now multiply odd parts.
15059 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
15061 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
15062 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
15064 // Merge the two vectors back together with a shuffle. This expands into 2
15066 static const int ShufMask[] = { 0, 4, 2, 6 };
15067 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
15070 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
15071 "Only know how to lower V2I64/V4I64/V8I64 multiply");
15073 // Ahi = psrlqi(a, 32);
15074 // Bhi = psrlqi(b, 32);
15076 // AloBlo = pmuludq(a, b);
15077 // AloBhi = pmuludq(a, Bhi);
15078 // AhiBlo = pmuludq(Ahi, b);
15080 // AloBhi = psllqi(AloBhi, 32);
15081 // AhiBlo = psllqi(AhiBlo, 32);
15082 // return AloBlo + AloBhi + AhiBlo;
15084 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
15085 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
15087 // Bit cast to 32-bit vectors for MULUDQ
15088 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
15089 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
15090 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
15091 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
15092 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
15093 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
15095 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
15096 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
15097 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
15099 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
15100 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
15102 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
15103 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
15106 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
15107 assert(Subtarget->isTargetWin64() && "Unexpected target");
15108 EVT VT = Op.getValueType();
15109 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
15110 "Unexpected return type for lowering");
15114 switch (Op->getOpcode()) {
15115 default: llvm_unreachable("Unexpected request for libcall!");
15116 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
15117 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
15118 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
15119 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
15120 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
15121 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
15125 SDValue InChain = DAG.getEntryNode();
15127 TargetLowering::ArgListTy Args;
15128 TargetLowering::ArgListEntry Entry;
15129 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
15130 EVT ArgVT = Op->getOperand(i).getValueType();
15131 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
15132 "Unexpected argument type for lowering");
15133 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
15134 Entry.Node = StackPtr;
15135 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
15137 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15138 Entry.Ty = PointerType::get(ArgTy,0);
15139 Entry.isSExt = false;
15140 Entry.isZExt = false;
15141 Args.push_back(Entry);
15144 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
15147 TargetLowering::CallLoweringInfo CLI(DAG);
15148 CLI.setDebugLoc(dl).setChain(InChain)
15149 .setCallee(getLibcallCallingConv(LC),
15150 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
15151 Callee, std::move(Args), 0)
15152 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
15154 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
15155 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
15158 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
15159 SelectionDAG &DAG) {
15160 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
15161 EVT VT = Op0.getValueType();
15164 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
15165 (VT == MVT::v8i32 && Subtarget->hasInt256()));
15167 // PMULxD operations multiply each even value (starting at 0) of LHS with
15168 // the related value of RHS and produce a widen result.
15169 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
15170 // => <2 x i64> <ae|cg>
15172 // In other word, to have all the results, we need to perform two PMULxD:
15173 // 1. one with the even values.
15174 // 2. one with the odd values.
15175 // To achieve #2, with need to place the odd values at an even position.
15177 // Place the odd value at an even position (basically, shift all values 1
15178 // step to the left):
15179 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
15180 // <a|b|c|d> => <b|undef|d|undef>
15181 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
15182 // <e|f|g|h> => <f|undef|h|undef>
15183 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
15185 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
15187 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
15188 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
15190 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
15191 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
15192 // => <2 x i64> <ae|cg>
15193 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
15194 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
15195 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
15196 // => <2 x i64> <bf|dh>
15197 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
15198 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
15200 // Shuffle it back into the right order.
15201 // The internal representation is big endian.
15202 // In other words, a i64 bitcasted to 2 x i32 has its high part at index 0
15203 // and its low part at index 1.
15204 // Moreover, we have: Mul1 = <ae|cg> ; Mul2 = <bf|dh>
15205 // Vector index 0 1 ; 2 3
15206 // We want <ae|bf|cg|dh>
15207 // Vector index 0 2 1 3
15208 // Since each element is seen as 2 x i32, we get:
15209 // high_mask[i] = 2 x vector_index[i]
15210 // low_mask[i] = 2 x vector_index[i] + 1
15211 // where vector_index = {0, Size/2, 1, Size/2 + 1, ...,
15212 // Size/2 - 1, Size/2 + Size/2 - 1}
15213 // where Size is the number of element of the final vector.
15214 SDValue Highs, Lows;
15215 if (VT == MVT::v8i32) {
15216 const int HighMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
15217 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
15218 const int LowMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
15219 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
15221 const int HighMask[] = {0, 4, 2, 6};
15222 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
15223 const int LowMask[] = {1, 5, 3, 7};
15224 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
15227 // If we have a signed multiply but no PMULDQ fix up the high parts of a
15228 // unsigned multiply.
15229 if (IsSigned && !Subtarget->hasSSE41()) {
15231 DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
15232 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
15233 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
15234 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
15235 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
15237 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
15238 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
15241 // The low part of a MUL_LOHI is supposed to be the first value and the
15242 // high part the second value.
15243 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getValueType(), Lows, Highs);
15246 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
15247 const X86Subtarget *Subtarget) {
15248 MVT VT = Op.getSimpleValueType();
15250 SDValue R = Op.getOperand(0);
15251 SDValue Amt = Op.getOperand(1);
15253 // Optimize shl/srl/sra with constant shift amount.
15254 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
15255 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
15256 uint64_t ShiftAmt = ShiftConst->getZExtValue();
15258 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
15259 (Subtarget->hasInt256() &&
15260 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
15261 (Subtarget->hasAVX512() &&
15262 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
15263 if (Op.getOpcode() == ISD::SHL)
15264 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
15266 if (Op.getOpcode() == ISD::SRL)
15267 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
15269 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
15270 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
15274 if (VT == MVT::v16i8) {
15275 if (Op.getOpcode() == ISD::SHL) {
15276 // Make a large shift.
15277 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
15278 MVT::v8i16, R, ShiftAmt,
15280 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
15281 // Zero out the rightmost bits.
15282 SmallVector<SDValue, 16> V(16,
15283 DAG.getConstant(uint8_t(-1U << ShiftAmt),
15285 return DAG.getNode(ISD::AND, dl, VT, SHL,
15286 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15288 if (Op.getOpcode() == ISD::SRL) {
15289 // Make a large shift.
15290 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
15291 MVT::v8i16, R, ShiftAmt,
15293 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
15294 // Zero out the leftmost bits.
15295 SmallVector<SDValue, 16> V(16,
15296 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
15298 return DAG.getNode(ISD::AND, dl, VT, SRL,
15299 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15301 if (Op.getOpcode() == ISD::SRA) {
15302 if (ShiftAmt == 7) {
15303 // R s>> 7 === R s< 0
15304 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
15305 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
15308 // R s>> a === ((R u>> a) ^ m) - m
15309 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
15310 SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
15312 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
15313 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
15314 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
15317 llvm_unreachable("Unknown shift opcode.");
15320 if (Subtarget->hasInt256() && VT == MVT::v32i8) {
15321 if (Op.getOpcode() == ISD::SHL) {
15322 // Make a large shift.
15323 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
15324 MVT::v16i16, R, ShiftAmt,
15326 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
15327 // Zero out the rightmost bits.
15328 SmallVector<SDValue, 32> V(32,
15329 DAG.getConstant(uint8_t(-1U << ShiftAmt),
15331 return DAG.getNode(ISD::AND, dl, VT, SHL,
15332 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15334 if (Op.getOpcode() == ISD::SRL) {
15335 // Make a large shift.
15336 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
15337 MVT::v16i16, R, ShiftAmt,
15339 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
15340 // Zero out the leftmost bits.
15341 SmallVector<SDValue, 32> V(32,
15342 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
15344 return DAG.getNode(ISD::AND, dl, VT, SRL,
15345 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15347 if (Op.getOpcode() == ISD::SRA) {
15348 if (ShiftAmt == 7) {
15349 // R s>> 7 === R s< 0
15350 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
15351 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
15354 // R s>> a === ((R u>> a) ^ m) - m
15355 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
15356 SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
15358 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
15359 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
15360 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
15363 llvm_unreachable("Unknown shift opcode.");
15368 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
15369 if (!Subtarget->is64Bit() &&
15370 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
15371 Amt.getOpcode() == ISD::BITCAST &&
15372 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
15373 Amt = Amt.getOperand(0);
15374 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
15375 VT.getVectorNumElements();
15376 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
15377 uint64_t ShiftAmt = 0;
15378 for (unsigned i = 0; i != Ratio; ++i) {
15379 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
15383 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
15385 // Check remaining shift amounts.
15386 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
15387 uint64_t ShAmt = 0;
15388 for (unsigned j = 0; j != Ratio; ++j) {
15389 ConstantSDNode *C =
15390 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
15394 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
15396 if (ShAmt != ShiftAmt)
15399 switch (Op.getOpcode()) {
15401 llvm_unreachable("Unknown shift opcode!");
15403 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
15406 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
15409 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
15417 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
15418 const X86Subtarget* Subtarget) {
15419 MVT VT = Op.getSimpleValueType();
15421 SDValue R = Op.getOperand(0);
15422 SDValue Amt = Op.getOperand(1);
15424 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
15425 VT == MVT::v4i32 || VT == MVT::v8i16 ||
15426 (Subtarget->hasInt256() &&
15427 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
15428 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
15429 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
15431 EVT EltVT = VT.getVectorElementType();
15433 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
15434 unsigned NumElts = VT.getVectorNumElements();
15436 for (i = 0; i != NumElts; ++i) {
15437 if (Amt.getOperand(i).getOpcode() == ISD::UNDEF)
15441 for (j = i; j != NumElts; ++j) {
15442 SDValue Arg = Amt.getOperand(j);
15443 if (Arg.getOpcode() == ISD::UNDEF) continue;
15444 if (Arg != Amt.getOperand(i))
15447 if (i != NumElts && j == NumElts)
15448 BaseShAmt = Amt.getOperand(i);
15450 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
15451 Amt = Amt.getOperand(0);
15452 if (Amt.getOpcode() == ISD::VECTOR_SHUFFLE &&
15453 cast<ShuffleVectorSDNode>(Amt)->isSplat()) {
15454 SDValue InVec = Amt.getOperand(0);
15455 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
15456 unsigned NumElts = InVec.getValueType().getVectorNumElements();
15458 for (; i != NumElts; ++i) {
15459 SDValue Arg = InVec.getOperand(i);
15460 if (Arg.getOpcode() == ISD::UNDEF) continue;
15464 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
15465 if (ConstantSDNode *C =
15466 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
15467 unsigned SplatIdx =
15468 cast<ShuffleVectorSDNode>(Amt)->getSplatIndex();
15469 if (C->getZExtValue() == SplatIdx)
15470 BaseShAmt = InVec.getOperand(1);
15473 if (!BaseShAmt.getNode())
15474 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Amt,
15475 DAG.getIntPtrConstant(0));
15479 if (BaseShAmt.getNode()) {
15480 if (EltVT.bitsGT(MVT::i32))
15481 BaseShAmt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BaseShAmt);
15482 else if (EltVT.bitsLT(MVT::i32))
15483 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
15485 switch (Op.getOpcode()) {
15487 llvm_unreachable("Unknown shift opcode!");
15489 switch (VT.SimpleTy) {
15490 default: return SDValue();
15499 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
15502 switch (VT.SimpleTy) {
15503 default: return SDValue();
15510 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
15513 switch (VT.SimpleTy) {
15514 default: return SDValue();
15523 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
15529 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
15530 if (!Subtarget->is64Bit() &&
15531 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
15532 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
15533 Amt.getOpcode() == ISD::BITCAST &&
15534 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
15535 Amt = Amt.getOperand(0);
15536 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
15537 VT.getVectorNumElements();
15538 std::vector<SDValue> Vals(Ratio);
15539 for (unsigned i = 0; i != Ratio; ++i)
15540 Vals[i] = Amt.getOperand(i);
15541 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
15542 for (unsigned j = 0; j != Ratio; ++j)
15543 if (Vals[j] != Amt.getOperand(i + j))
15546 switch (Op.getOpcode()) {
15548 llvm_unreachable("Unknown shift opcode!");
15550 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
15552 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
15554 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
15561 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
15562 SelectionDAG &DAG) {
15563 MVT VT = Op.getSimpleValueType();
15565 SDValue R = Op.getOperand(0);
15566 SDValue Amt = Op.getOperand(1);
15569 assert(VT.isVector() && "Custom lowering only for vector shifts!");
15570 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
15572 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
15576 V = LowerScalarVariableShift(Op, DAG, Subtarget);
15580 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
15582 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
15583 if (Subtarget->hasInt256()) {
15584 if (Op.getOpcode() == ISD::SRL &&
15585 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
15586 VT == MVT::v4i64 || VT == MVT::v8i32))
15588 if (Op.getOpcode() == ISD::SHL &&
15589 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
15590 VT == MVT::v4i64 || VT == MVT::v8i32))
15592 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
15596 // If possible, lower this packed shift into a vector multiply instead of
15597 // expanding it into a sequence of scalar shifts.
15598 // Do this only if the vector shift count is a constant build_vector.
15599 if (Op.getOpcode() == ISD::SHL &&
15600 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
15601 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
15602 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
15603 SmallVector<SDValue, 8> Elts;
15604 EVT SVT = VT.getScalarType();
15605 unsigned SVTBits = SVT.getSizeInBits();
15606 const APInt &One = APInt(SVTBits, 1);
15607 unsigned NumElems = VT.getVectorNumElements();
15609 for (unsigned i=0; i !=NumElems; ++i) {
15610 SDValue Op = Amt->getOperand(i);
15611 if (Op->getOpcode() == ISD::UNDEF) {
15612 Elts.push_back(Op);
15616 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
15617 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
15618 uint64_t ShAmt = C.getZExtValue();
15619 if (ShAmt >= SVTBits) {
15620 Elts.push_back(DAG.getUNDEF(SVT));
15623 Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
15625 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15626 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
15629 // Lower SHL with variable shift amount.
15630 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
15631 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
15633 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
15634 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
15635 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
15636 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
15639 // If possible, lower this shift as a sequence of two shifts by
15640 // constant plus a MOVSS/MOVSD instead of scalarizing it.
15642 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
15644 // Could be rewritten as:
15645 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
15647 // The advantage is that the two shifts from the example would be
15648 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
15649 // the vector shift into four scalar shifts plus four pairs of vector
15651 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
15652 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
15653 unsigned TargetOpcode = X86ISD::MOVSS;
15654 bool CanBeSimplified;
15655 // The splat value for the first packed shift (the 'X' from the example).
15656 SDValue Amt1 = Amt->getOperand(0);
15657 // The splat value for the second packed shift (the 'Y' from the example).
15658 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
15659 Amt->getOperand(2);
15661 // See if it is possible to replace this node with a sequence of
15662 // two shifts followed by a MOVSS/MOVSD
15663 if (VT == MVT::v4i32) {
15664 // Check if it is legal to use a MOVSS.
15665 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
15666 Amt2 == Amt->getOperand(3);
15667 if (!CanBeSimplified) {
15668 // Otherwise, check if we can still simplify this node using a MOVSD.
15669 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
15670 Amt->getOperand(2) == Amt->getOperand(3);
15671 TargetOpcode = X86ISD::MOVSD;
15672 Amt2 = Amt->getOperand(2);
15675 // Do similar checks for the case where the machine value type
15677 CanBeSimplified = Amt1 == Amt->getOperand(1);
15678 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
15679 CanBeSimplified = Amt2 == Amt->getOperand(i);
15681 if (!CanBeSimplified) {
15682 TargetOpcode = X86ISD::MOVSD;
15683 CanBeSimplified = true;
15684 Amt2 = Amt->getOperand(4);
15685 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
15686 CanBeSimplified = Amt1 == Amt->getOperand(i);
15687 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
15688 CanBeSimplified = Amt2 == Amt->getOperand(j);
15692 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
15693 isa<ConstantSDNode>(Amt2)) {
15694 // Replace this node with two shifts followed by a MOVSS/MOVSD.
15695 EVT CastVT = MVT::v4i32;
15697 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
15698 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
15700 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
15701 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
15702 if (TargetOpcode == X86ISD::MOVSD)
15703 CastVT = MVT::v2i64;
15704 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
15705 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
15706 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
15708 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
15712 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
15713 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
15716 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
15717 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
15719 // Turn 'a' into a mask suitable for VSELECT
15720 SDValue VSelM = DAG.getConstant(0x80, VT);
15721 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
15722 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
15724 SDValue CM1 = DAG.getConstant(0x0f, VT);
15725 SDValue CM2 = DAG.getConstant(0x3f, VT);
15727 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
15728 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
15729 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG);
15730 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
15731 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
15734 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
15735 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
15736 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
15738 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
15739 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
15740 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG);
15741 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
15742 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
15745 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
15746 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
15747 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
15749 // return VSELECT(r, r+r, a);
15750 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
15751 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
15755 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
15756 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
15757 // solution better.
15758 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
15759 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
15761 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
15762 R = DAG.getNode(ExtOpc, dl, NewVT, R);
15763 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
15764 return DAG.getNode(ISD::TRUNCATE, dl, VT,
15765 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
15768 // Decompose 256-bit shifts into smaller 128-bit shifts.
15769 if (VT.is256BitVector()) {
15770 unsigned NumElems = VT.getVectorNumElements();
15771 MVT EltVT = VT.getVectorElementType();
15772 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15774 // Extract the two vectors
15775 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
15776 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
15778 // Recreate the shift amount vectors
15779 SDValue Amt1, Amt2;
15780 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
15781 // Constant shift amount
15782 SmallVector<SDValue, 4> Amt1Csts;
15783 SmallVector<SDValue, 4> Amt2Csts;
15784 for (unsigned i = 0; i != NumElems/2; ++i)
15785 Amt1Csts.push_back(Amt->getOperand(i));
15786 for (unsigned i = NumElems/2; i != NumElems; ++i)
15787 Amt2Csts.push_back(Amt->getOperand(i));
15789 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
15790 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
15792 // Variable shift amount
15793 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
15794 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
15797 // Issue new vector shifts for the smaller types
15798 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
15799 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
15801 // Concatenate the result back
15802 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
15808 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
15809 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
15810 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
15811 // looks for this combo and may remove the "setcc" instruction if the "setcc"
15812 // has only one use.
15813 SDNode *N = Op.getNode();
15814 SDValue LHS = N->getOperand(0);
15815 SDValue RHS = N->getOperand(1);
15816 unsigned BaseOp = 0;
15819 switch (Op.getOpcode()) {
15820 default: llvm_unreachable("Unknown ovf instruction!");
15822 // A subtract of one will be selected as a INC. Note that INC doesn't
15823 // set CF, so we can't do this for UADDO.
15824 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15826 BaseOp = X86ISD::INC;
15827 Cond = X86::COND_O;
15830 BaseOp = X86ISD::ADD;
15831 Cond = X86::COND_O;
15834 BaseOp = X86ISD::ADD;
15835 Cond = X86::COND_B;
15838 // A subtract of one will be selected as a DEC. Note that DEC doesn't
15839 // set CF, so we can't do this for USUBO.
15840 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15842 BaseOp = X86ISD::DEC;
15843 Cond = X86::COND_O;
15846 BaseOp = X86ISD::SUB;
15847 Cond = X86::COND_O;
15850 BaseOp = X86ISD::SUB;
15851 Cond = X86::COND_B;
15854 BaseOp = X86ISD::SMUL;
15855 Cond = X86::COND_O;
15857 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
15858 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
15860 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
15863 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
15864 DAG.getConstant(X86::COND_O, MVT::i32),
15865 SDValue(Sum.getNode(), 2));
15867 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
15871 // Also sets EFLAGS.
15872 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
15873 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
15876 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
15877 DAG.getConstant(Cond, MVT::i32),
15878 SDValue(Sum.getNode(), 1));
15880 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
15883 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
15884 SelectionDAG &DAG) const {
15886 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
15887 MVT VT = Op.getSimpleValueType();
15889 if (!Subtarget->hasSSE2() || !VT.isVector())
15892 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
15893 ExtraVT.getScalarType().getSizeInBits();
15895 switch (VT.SimpleTy) {
15896 default: return SDValue();
15899 if (!Subtarget->hasFp256())
15901 if (!Subtarget->hasInt256()) {
15902 // needs to be split
15903 unsigned NumElems = VT.getVectorNumElements();
15905 // Extract the LHS vectors
15906 SDValue LHS = Op.getOperand(0);
15907 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15908 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15910 MVT EltVT = VT.getVectorElementType();
15911 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15913 EVT ExtraEltVT = ExtraVT.getVectorElementType();
15914 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
15915 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
15917 SDValue Extra = DAG.getValueType(ExtraVT);
15919 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
15920 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
15922 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
15927 SDValue Op0 = Op.getOperand(0);
15928 SDValue Op00 = Op0.getOperand(0);
15930 // Hopefully, this VECTOR_SHUFFLE is just a VZEXT.
15931 if (Op0.getOpcode() == ISD::BITCAST &&
15932 Op00.getOpcode() == ISD::VECTOR_SHUFFLE) {
15933 // (sext (vzext x)) -> (vsext x)
15934 Tmp1 = LowerVectorIntExtend(Op00, Subtarget, DAG);
15935 if (Tmp1.getNode()) {
15936 EVT ExtraEltVT = ExtraVT.getVectorElementType();
15937 // This folding is only valid when the in-reg type is a vector of i8,
15939 if (ExtraEltVT == MVT::i8 || ExtraEltVT == MVT::i16 ||
15940 ExtraEltVT == MVT::i32) {
15941 SDValue Tmp1Op0 = Tmp1.getOperand(0);
15942 assert(Tmp1Op0.getOpcode() == X86ISD::VZEXT &&
15943 "This optimization is invalid without a VZEXT.");
15944 return DAG.getNode(X86ISD::VSEXT, dl, VT, Tmp1Op0.getOperand(0));
15950 // If the above didn't work, then just use Shift-Left + Shift-Right.
15951 Tmp1 = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0, BitsDiff,
15953 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Tmp1, BitsDiff,
15959 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
15960 SelectionDAG &DAG) {
15962 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
15963 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
15964 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
15965 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
15967 // The only fence that needs an instruction is a sequentially-consistent
15968 // cross-thread fence.
15969 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
15970 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
15971 // no-sse2). There isn't any reason to disable it if the target processor
15973 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
15974 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
15976 SDValue Chain = Op.getOperand(0);
15977 SDValue Zero = DAG.getConstant(0, MVT::i32);
15979 DAG.getRegister(X86::ESP, MVT::i32), // Base
15980 DAG.getTargetConstant(1, MVT::i8), // Scale
15981 DAG.getRegister(0, MVT::i32), // Index
15982 DAG.getTargetConstant(0, MVT::i32), // Disp
15983 DAG.getRegister(0, MVT::i32), // Segment.
15987 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
15988 return SDValue(Res, 0);
15991 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
15992 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
15995 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
15996 SelectionDAG &DAG) {
15997 MVT T = Op.getSimpleValueType();
16001 switch(T.SimpleTy) {
16002 default: llvm_unreachable("Invalid value type!");
16003 case MVT::i8: Reg = X86::AL; size = 1; break;
16004 case MVT::i16: Reg = X86::AX; size = 2; break;
16005 case MVT::i32: Reg = X86::EAX; size = 4; break;
16007 assert(Subtarget->is64Bit() && "Node not type legal!");
16008 Reg = X86::RAX; size = 8;
16011 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
16012 Op.getOperand(2), SDValue());
16013 SDValue Ops[] = { cpIn.getValue(0),
16016 DAG.getTargetConstant(size, MVT::i8),
16017 cpIn.getValue(1) };
16018 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16019 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
16020 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
16024 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
16025 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
16026 MVT::i32, cpOut.getValue(2));
16027 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
16028 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
16030 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
16031 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
16032 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
16036 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
16037 SelectionDAG &DAG) {
16038 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
16039 MVT DstVT = Op.getSimpleValueType();
16041 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
16042 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
16043 if (DstVT != MVT::f64)
16044 // This conversion needs to be expanded.
16047 SDValue InVec = Op->getOperand(0);
16049 unsigned NumElts = SrcVT.getVectorNumElements();
16050 EVT SVT = SrcVT.getVectorElementType();
16052 // Widen the vector in input in the case of MVT::v2i32.
16053 // Example: from MVT::v2i32 to MVT::v4i32.
16054 SmallVector<SDValue, 16> Elts;
16055 for (unsigned i = 0, e = NumElts; i != e; ++i)
16056 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
16057 DAG.getIntPtrConstant(i)));
16059 // Explicitly mark the extra elements as Undef.
16060 SDValue Undef = DAG.getUNDEF(SVT);
16061 for (unsigned i = NumElts, e = NumElts * 2; i != e; ++i)
16062 Elts.push_back(Undef);
16064 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
16065 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
16066 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
16067 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
16068 DAG.getIntPtrConstant(0));
16071 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
16072 Subtarget->hasMMX() && "Unexpected custom BITCAST");
16073 assert((DstVT == MVT::i64 ||
16074 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
16075 "Unexpected custom BITCAST");
16076 // i64 <=> MMX conversions are Legal.
16077 if (SrcVT==MVT::i64 && DstVT.isVector())
16079 if (DstVT==MVT::i64 && SrcVT.isVector())
16081 // MMX <=> MMX conversions are Legal.
16082 if (SrcVT.isVector() && DstVT.isVector())
16084 // All other conversions need to be expanded.
16088 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
16089 SDNode *Node = Op.getNode();
16091 EVT T = Node->getValueType(0);
16092 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
16093 DAG.getConstant(0, T), Node->getOperand(2));
16094 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
16095 cast<AtomicSDNode>(Node)->getMemoryVT(),
16096 Node->getOperand(0),
16097 Node->getOperand(1), negOp,
16098 cast<AtomicSDNode>(Node)->getMemOperand(),
16099 cast<AtomicSDNode>(Node)->getOrdering(),
16100 cast<AtomicSDNode>(Node)->getSynchScope());
16103 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
16104 SDNode *Node = Op.getNode();
16106 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
16108 // Convert seq_cst store -> xchg
16109 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
16110 // FIXME: On 32-bit, store -> fist or movq would be more efficient
16111 // (The only way to get a 16-byte store is cmpxchg16b)
16112 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
16113 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
16114 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
16115 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
16116 cast<AtomicSDNode>(Node)->getMemoryVT(),
16117 Node->getOperand(0),
16118 Node->getOperand(1), Node->getOperand(2),
16119 cast<AtomicSDNode>(Node)->getMemOperand(),
16120 cast<AtomicSDNode>(Node)->getOrdering(),
16121 cast<AtomicSDNode>(Node)->getSynchScope());
16122 return Swap.getValue(1);
16124 // Other atomic stores have a simple pattern.
16128 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
16129 EVT VT = Op.getNode()->getSimpleValueType(0);
16131 // Let legalize expand this if it isn't a legal type yet.
16132 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
16135 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16138 bool ExtraOp = false;
16139 switch (Op.getOpcode()) {
16140 default: llvm_unreachable("Invalid code");
16141 case ISD::ADDC: Opc = X86ISD::ADD; break;
16142 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
16143 case ISD::SUBC: Opc = X86ISD::SUB; break;
16144 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
16148 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
16150 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
16151 Op.getOperand(1), Op.getOperand(2));
16154 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
16155 SelectionDAG &DAG) {
16156 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
16158 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
16159 // which returns the values as { float, float } (in XMM0) or
16160 // { double, double } (which is returned in XMM0, XMM1).
16162 SDValue Arg = Op.getOperand(0);
16163 EVT ArgVT = Arg.getValueType();
16164 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16166 TargetLowering::ArgListTy Args;
16167 TargetLowering::ArgListEntry Entry;
16171 Entry.isSExt = false;
16172 Entry.isZExt = false;
16173 Args.push_back(Entry);
16175 bool isF64 = ArgVT == MVT::f64;
16176 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
16177 // the small struct {f32, f32} is returned in (eax, edx). For f64,
16178 // the results are returned via SRet in memory.
16179 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
16180 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16181 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
16183 Type *RetTy = isF64
16184 ? (Type*)StructType::get(ArgTy, ArgTy, NULL)
16185 : (Type*)VectorType::get(ArgTy, 4);
16187 TargetLowering::CallLoweringInfo CLI(DAG);
16188 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
16189 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
16191 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
16194 // Returned in xmm0 and xmm1.
16195 return CallResult.first;
16197 // Returned in bits 0:31 and 32:64 xmm0.
16198 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
16199 CallResult.first, DAG.getIntPtrConstant(0));
16200 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
16201 CallResult.first, DAG.getIntPtrConstant(1));
16202 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
16203 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
16206 /// LowerOperation - Provide custom lowering hooks for some operations.
16208 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
16209 switch (Op.getOpcode()) {
16210 default: llvm_unreachable("Should not custom lower this!");
16211 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
16212 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
16213 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
16214 return LowerCMP_SWAP(Op, Subtarget, DAG);
16215 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
16216 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
16217 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
16218 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
16219 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
16220 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
16221 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
16222 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
16223 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
16224 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
16225 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
16226 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
16227 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
16228 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
16229 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
16230 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
16231 case ISD::SHL_PARTS:
16232 case ISD::SRA_PARTS:
16233 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
16234 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
16235 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
16236 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
16237 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
16238 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
16239 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
16240 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
16241 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
16242 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
16243 case ISD::FABS: return LowerFABS(Op, DAG);
16244 case ISD::FNEG: return LowerFNEG(Op, DAG);
16245 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
16246 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
16247 case ISD::SETCC: return LowerSETCC(Op, DAG);
16248 case ISD::SELECT: return LowerSELECT(Op, DAG);
16249 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
16250 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
16251 case ISD::VASTART: return LowerVASTART(Op, DAG);
16252 case ISD::VAARG: return LowerVAARG(Op, DAG);
16253 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
16254 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
16255 case ISD::INTRINSIC_VOID:
16256 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
16257 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
16258 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
16259 case ISD::FRAME_TO_ARGS_OFFSET:
16260 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
16261 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
16262 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
16263 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
16264 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
16265 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
16266 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
16267 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
16268 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
16269 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
16270 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
16271 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
16272 case ISD::UMUL_LOHI:
16273 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
16276 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
16282 case ISD::UMULO: return LowerXALUO(Op, DAG);
16283 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
16284 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
16288 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
16289 case ISD::ADD: return LowerADD(Op, DAG);
16290 case ISD::SUB: return LowerSUB(Op, DAG);
16291 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
16295 static void ReplaceATOMIC_LOAD(SDNode *Node,
16296 SmallVectorImpl<SDValue> &Results,
16297 SelectionDAG &DAG) {
16299 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
16301 // Convert wide load -> cmpxchg8b/cmpxchg16b
16302 // FIXME: On 32-bit, load -> fild or movq would be more efficient
16303 // (The only way to get a 16-byte load is cmpxchg16b)
16304 // FIXME: 16-byte ATOMIC_CMP_SWAP isn't actually hooked up at the moment.
16305 SDValue Zero = DAG.getConstant(0, VT);
16306 SDVTList VTs = DAG.getVTList(VT, MVT::i1, MVT::Other);
16308 DAG.getAtomicCmpSwap(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, dl, VT, VTs,
16309 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
16310 cast<AtomicSDNode>(Node)->getMemOperand(),
16311 cast<AtomicSDNode>(Node)->getOrdering(),
16312 cast<AtomicSDNode>(Node)->getOrdering(),
16313 cast<AtomicSDNode>(Node)->getSynchScope());
16314 Results.push_back(Swap.getValue(0));
16315 Results.push_back(Swap.getValue(2));
16318 /// ReplaceNodeResults - Replace a node with an illegal result type
16319 /// with a new node built out of custom code.
16320 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
16321 SmallVectorImpl<SDValue>&Results,
16322 SelectionDAG &DAG) const {
16324 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16325 switch (N->getOpcode()) {
16327 llvm_unreachable("Do not know how to custom type legalize this operation!");
16328 case ISD::SIGN_EXTEND_INREG:
16333 // We don't want to expand or promote these.
16340 case ISD::UDIVREM: {
16341 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
16342 Results.push_back(V);
16345 case ISD::FP_TO_SINT:
16346 case ISD::FP_TO_UINT: {
16347 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
16349 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
16352 std::pair<SDValue,SDValue> Vals =
16353 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
16354 SDValue FIST = Vals.first, StackSlot = Vals.second;
16355 if (FIST.getNode()) {
16356 EVT VT = N->getValueType(0);
16357 // Return a load from the stack slot.
16358 if (StackSlot.getNode())
16359 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
16360 MachinePointerInfo(),
16361 false, false, false, 0));
16363 Results.push_back(FIST);
16367 case ISD::UINT_TO_FP: {
16368 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
16369 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
16370 N->getValueType(0) != MVT::v2f32)
16372 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
16374 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
16376 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
16377 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
16378 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
16379 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
16380 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
16381 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
16384 case ISD::FP_ROUND: {
16385 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
16387 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
16388 Results.push_back(V);
16391 case ISD::INTRINSIC_W_CHAIN: {
16392 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
16394 default : llvm_unreachable("Do not know how to custom type "
16395 "legalize this intrinsic operation!");
16396 case Intrinsic::x86_rdtsc:
16397 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
16399 case Intrinsic::x86_rdtscp:
16400 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
16402 case Intrinsic::x86_rdpmc:
16403 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
16406 case ISD::READCYCLECOUNTER: {
16407 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
16410 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
16411 EVT T = N->getValueType(0);
16412 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
16413 bool Regs64bit = T == MVT::i128;
16414 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
16415 SDValue cpInL, cpInH;
16416 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
16417 DAG.getConstant(0, HalfT));
16418 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
16419 DAG.getConstant(1, HalfT));
16420 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
16421 Regs64bit ? X86::RAX : X86::EAX,
16423 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
16424 Regs64bit ? X86::RDX : X86::EDX,
16425 cpInH, cpInL.getValue(1));
16426 SDValue swapInL, swapInH;
16427 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
16428 DAG.getConstant(0, HalfT));
16429 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
16430 DAG.getConstant(1, HalfT));
16431 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
16432 Regs64bit ? X86::RBX : X86::EBX,
16433 swapInL, cpInH.getValue(1));
16434 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
16435 Regs64bit ? X86::RCX : X86::ECX,
16436 swapInH, swapInL.getValue(1));
16437 SDValue Ops[] = { swapInH.getValue(0),
16439 swapInH.getValue(1) };
16440 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16441 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
16442 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
16443 X86ISD::LCMPXCHG8_DAG;
16444 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
16445 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
16446 Regs64bit ? X86::RAX : X86::EAX,
16447 HalfT, Result.getValue(1));
16448 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
16449 Regs64bit ? X86::RDX : X86::EDX,
16450 HalfT, cpOutL.getValue(2));
16451 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
16453 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
16454 MVT::i32, cpOutH.getValue(2));
16456 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16457 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
16458 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
16460 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
16461 Results.push_back(Success);
16462 Results.push_back(EFLAGS.getValue(1));
16465 case ISD::ATOMIC_SWAP:
16466 case ISD::ATOMIC_LOAD_ADD:
16467 case ISD::ATOMIC_LOAD_SUB:
16468 case ISD::ATOMIC_LOAD_AND:
16469 case ISD::ATOMIC_LOAD_OR:
16470 case ISD::ATOMIC_LOAD_XOR:
16471 case ISD::ATOMIC_LOAD_NAND:
16472 case ISD::ATOMIC_LOAD_MIN:
16473 case ISD::ATOMIC_LOAD_MAX:
16474 case ISD::ATOMIC_LOAD_UMIN:
16475 case ISD::ATOMIC_LOAD_UMAX:
16476 // Delegate to generic TypeLegalization. Situations we can really handle
16477 // should have already been dealt with by X86AtomicExpand.cpp.
16479 case ISD::ATOMIC_LOAD: {
16480 ReplaceATOMIC_LOAD(N, Results, DAG);
16483 case ISD::BITCAST: {
16484 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
16485 EVT DstVT = N->getValueType(0);
16486 EVT SrcVT = N->getOperand(0)->getValueType(0);
16488 if (SrcVT != MVT::f64 ||
16489 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
16492 unsigned NumElts = DstVT.getVectorNumElements();
16493 EVT SVT = DstVT.getVectorElementType();
16494 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
16495 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
16496 MVT::v2f64, N->getOperand(0));
16497 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
16499 if (ExperimentalVectorWideningLegalization) {
16500 // If we are legalizing vectors by widening, we already have the desired
16501 // legal vector type, just return it.
16502 Results.push_back(ToVecInt);
16506 SmallVector<SDValue, 8> Elts;
16507 for (unsigned i = 0, e = NumElts; i != e; ++i)
16508 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
16509 ToVecInt, DAG.getIntPtrConstant(i)));
16511 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
16516 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
16518 default: return nullptr;
16519 case X86ISD::BSF: return "X86ISD::BSF";
16520 case X86ISD::BSR: return "X86ISD::BSR";
16521 case X86ISD::SHLD: return "X86ISD::SHLD";
16522 case X86ISD::SHRD: return "X86ISD::SHRD";
16523 case X86ISD::FAND: return "X86ISD::FAND";
16524 case X86ISD::FANDN: return "X86ISD::FANDN";
16525 case X86ISD::FOR: return "X86ISD::FOR";
16526 case X86ISD::FXOR: return "X86ISD::FXOR";
16527 case X86ISD::FSRL: return "X86ISD::FSRL";
16528 case X86ISD::FILD: return "X86ISD::FILD";
16529 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
16530 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
16531 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
16532 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
16533 case X86ISD::FLD: return "X86ISD::FLD";
16534 case X86ISD::FST: return "X86ISD::FST";
16535 case X86ISD::CALL: return "X86ISD::CALL";
16536 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
16537 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
16538 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
16539 case X86ISD::BT: return "X86ISD::BT";
16540 case X86ISD::CMP: return "X86ISD::CMP";
16541 case X86ISD::COMI: return "X86ISD::COMI";
16542 case X86ISD::UCOMI: return "X86ISD::UCOMI";
16543 case X86ISD::CMPM: return "X86ISD::CMPM";
16544 case X86ISD::CMPMU: return "X86ISD::CMPMU";
16545 case X86ISD::SETCC: return "X86ISD::SETCC";
16546 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
16547 case X86ISD::FSETCC: return "X86ISD::FSETCC";
16548 case X86ISD::CMOV: return "X86ISD::CMOV";
16549 case X86ISD::BRCOND: return "X86ISD::BRCOND";
16550 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
16551 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
16552 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
16553 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
16554 case X86ISD::Wrapper: return "X86ISD::Wrapper";
16555 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
16556 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
16557 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
16558 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
16559 case X86ISD::PINSRB: return "X86ISD::PINSRB";
16560 case X86ISD::PINSRW: return "X86ISD::PINSRW";
16561 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
16562 case X86ISD::ANDNP: return "X86ISD::ANDNP";
16563 case X86ISD::PSIGN: return "X86ISD::PSIGN";
16564 case X86ISD::BLENDV: return "X86ISD::BLENDV";
16565 case X86ISD::BLENDI: return "X86ISD::BLENDI";
16566 case X86ISD::SUBUS: return "X86ISD::SUBUS";
16567 case X86ISD::HADD: return "X86ISD::HADD";
16568 case X86ISD::HSUB: return "X86ISD::HSUB";
16569 case X86ISD::FHADD: return "X86ISD::FHADD";
16570 case X86ISD::FHSUB: return "X86ISD::FHSUB";
16571 case X86ISD::UMAX: return "X86ISD::UMAX";
16572 case X86ISD::UMIN: return "X86ISD::UMIN";
16573 case X86ISD::SMAX: return "X86ISD::SMAX";
16574 case X86ISD::SMIN: return "X86ISD::SMIN";
16575 case X86ISD::FMAX: return "X86ISD::FMAX";
16576 case X86ISD::FMIN: return "X86ISD::FMIN";
16577 case X86ISD::FMAXC: return "X86ISD::FMAXC";
16578 case X86ISD::FMINC: return "X86ISD::FMINC";
16579 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
16580 case X86ISD::FRCP: return "X86ISD::FRCP";
16581 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
16582 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
16583 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
16584 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
16585 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
16586 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
16587 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
16588 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
16589 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
16590 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
16591 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
16592 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
16593 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
16594 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
16595 case X86ISD::VZEXT: return "X86ISD::VZEXT";
16596 case X86ISD::VSEXT: return "X86ISD::VSEXT";
16597 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
16598 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
16599 case X86ISD::VINSERT: return "X86ISD::VINSERT";
16600 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
16601 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
16602 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
16603 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
16604 case X86ISD::VSHL: return "X86ISD::VSHL";
16605 case X86ISD::VSRL: return "X86ISD::VSRL";
16606 case X86ISD::VSRA: return "X86ISD::VSRA";
16607 case X86ISD::VSHLI: return "X86ISD::VSHLI";
16608 case X86ISD::VSRLI: return "X86ISD::VSRLI";
16609 case X86ISD::VSRAI: return "X86ISD::VSRAI";
16610 case X86ISD::CMPP: return "X86ISD::CMPP";
16611 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
16612 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
16613 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
16614 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
16615 case X86ISD::ADD: return "X86ISD::ADD";
16616 case X86ISD::SUB: return "X86ISD::SUB";
16617 case X86ISD::ADC: return "X86ISD::ADC";
16618 case X86ISD::SBB: return "X86ISD::SBB";
16619 case X86ISD::SMUL: return "X86ISD::SMUL";
16620 case X86ISD::UMUL: return "X86ISD::UMUL";
16621 case X86ISD::INC: return "X86ISD::INC";
16622 case X86ISD::DEC: return "X86ISD::DEC";
16623 case X86ISD::OR: return "X86ISD::OR";
16624 case X86ISD::XOR: return "X86ISD::XOR";
16625 case X86ISD::AND: return "X86ISD::AND";
16626 case X86ISD::BEXTR: return "X86ISD::BEXTR";
16627 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
16628 case X86ISD::PTEST: return "X86ISD::PTEST";
16629 case X86ISD::TESTP: return "X86ISD::TESTP";
16630 case X86ISD::TESTM: return "X86ISD::TESTM";
16631 case X86ISD::TESTNM: return "X86ISD::TESTNM";
16632 case X86ISD::KORTEST: return "X86ISD::KORTEST";
16633 case X86ISD::PACKSS: return "X86ISD::PACKSS";
16634 case X86ISD::PACKUS: return "X86ISD::PACKUS";
16635 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
16636 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
16637 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
16638 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
16639 case X86ISD::SHUFP: return "X86ISD::SHUFP";
16640 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
16641 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
16642 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
16643 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
16644 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
16645 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
16646 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
16647 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
16648 case X86ISD::MOVSD: return "X86ISD::MOVSD";
16649 case X86ISD::MOVSS: return "X86ISD::MOVSS";
16650 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
16651 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
16652 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
16653 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
16654 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
16655 case X86ISD::VPERMILP: return "X86ISD::VPERMILP";
16656 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
16657 case X86ISD::VPERMV: return "X86ISD::VPERMV";
16658 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
16659 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
16660 case X86ISD::VPERMI: return "X86ISD::VPERMI";
16661 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
16662 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
16663 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
16664 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
16665 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
16666 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
16667 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
16668 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
16669 case X86ISD::SAHF: return "X86ISD::SAHF";
16670 case X86ISD::RDRAND: return "X86ISD::RDRAND";
16671 case X86ISD::RDSEED: return "X86ISD::RDSEED";
16672 case X86ISD::FMADD: return "X86ISD::FMADD";
16673 case X86ISD::FMSUB: return "X86ISD::FMSUB";
16674 case X86ISD::FNMADD: return "X86ISD::FNMADD";
16675 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
16676 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
16677 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
16678 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
16679 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
16680 case X86ISD::XTEST: return "X86ISD::XTEST";
16684 // isLegalAddressingMode - Return true if the addressing mode represented
16685 // by AM is legal for this target, for a load/store of the specified type.
16686 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
16688 // X86 supports extremely general addressing modes.
16689 CodeModel::Model M = getTargetMachine().getCodeModel();
16690 Reloc::Model R = getTargetMachine().getRelocationModel();
16692 // X86 allows a sign-extended 32-bit immediate field as a displacement.
16693 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
16698 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
16700 // If a reference to this global requires an extra load, we can't fold it.
16701 if (isGlobalStubReference(GVFlags))
16704 // If BaseGV requires a register for the PIC base, we cannot also have a
16705 // BaseReg specified.
16706 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
16709 // If lower 4G is not available, then we must use rip-relative addressing.
16710 if ((M != CodeModel::Small || R != Reloc::Static) &&
16711 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
16715 switch (AM.Scale) {
16721 // These scales always work.
16726 // These scales are formed with basereg+scalereg. Only accept if there is
16731 default: // Other stuff never works.
16738 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
16739 unsigned Bits = Ty->getScalarSizeInBits();
16741 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
16742 // particularly cheaper than those without.
16746 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
16747 // variable shifts just as cheap as scalar ones.
16748 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
16751 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
16752 // fully general vector.
16756 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
16757 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
16759 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
16760 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
16761 return NumBits1 > NumBits2;
16764 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
16765 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
16768 if (!isTypeLegal(EVT::getEVT(Ty1)))
16771 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
16773 // Assuming the caller doesn't have a zeroext or signext return parameter,
16774 // truncation all the way down to i1 is valid.
16778 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
16779 return isInt<32>(Imm);
16782 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
16783 // Can also use sub to handle negated immediates.
16784 return isInt<32>(Imm);
16787 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
16788 if (!VT1.isInteger() || !VT2.isInteger())
16790 unsigned NumBits1 = VT1.getSizeInBits();
16791 unsigned NumBits2 = VT2.getSizeInBits();
16792 return NumBits1 > NumBits2;
16795 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
16796 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
16797 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
16800 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
16801 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
16802 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
16805 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
16806 EVT VT1 = Val.getValueType();
16807 if (isZExtFree(VT1, VT2))
16810 if (Val.getOpcode() != ISD::LOAD)
16813 if (!VT1.isSimple() || !VT1.isInteger() ||
16814 !VT2.isSimple() || !VT2.isInteger())
16817 switch (VT1.getSimpleVT().SimpleTy) {
16822 // X86 has 8, 16, and 32-bit zero-extending loads.
16830 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
16831 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
16834 VT = VT.getScalarType();
16836 if (!VT.isSimple())
16839 switch (VT.getSimpleVT().SimpleTy) {
16850 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
16851 // i16 instructions are longer (0x66 prefix) and potentially slower.
16852 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
16855 /// isShuffleMaskLegal - Targets can use this to indicate that they only
16856 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
16857 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
16858 /// are assumed to be legal.
16860 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
16862 if (!VT.isSimple())
16865 MVT SVT = VT.getSimpleVT();
16867 // Very little shuffling can be done for 64-bit vectors right now.
16868 if (VT.getSizeInBits() == 64)
16871 // If this is a single-input shuffle with no 128 bit lane crossings we can
16872 // lower it into pshufb.
16873 if ((SVT.is128BitVector() && Subtarget->hasSSSE3()) ||
16874 (SVT.is256BitVector() && Subtarget->hasInt256())) {
16875 bool isLegal = true;
16876 for (unsigned I = 0, E = M.size(); I != E; ++I) {
16877 if (M[I] >= (int)SVT.getVectorNumElements() ||
16878 ShuffleCrosses128bitLane(SVT, I, M[I])) {
16887 // FIXME: blends, shifts.
16888 return (SVT.getVectorNumElements() == 2 ||
16889 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
16890 isMOVLMask(M, SVT) ||
16891 isMOVHLPSMask(M, SVT) ||
16892 isSHUFPMask(M, SVT) ||
16893 isPSHUFDMask(M, SVT) ||
16894 isPSHUFHWMask(M, SVT, Subtarget->hasInt256()) ||
16895 isPSHUFLWMask(M, SVT, Subtarget->hasInt256()) ||
16896 isPALIGNRMask(M, SVT, Subtarget) ||
16897 isUNPCKLMask(M, SVT, Subtarget->hasInt256()) ||
16898 isUNPCKHMask(M, SVT, Subtarget->hasInt256()) ||
16899 isUNPCKL_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
16900 isUNPCKH_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
16901 isBlendMask(M, SVT, Subtarget->hasSSE41(), Subtarget->hasInt256()));
16905 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
16907 if (!VT.isSimple())
16910 MVT SVT = VT.getSimpleVT();
16911 unsigned NumElts = SVT.getVectorNumElements();
16912 // FIXME: This collection of masks seems suspect.
16915 if (NumElts == 4 && SVT.is128BitVector()) {
16916 return (isMOVLMask(Mask, SVT) ||
16917 isCommutedMOVLMask(Mask, SVT, true) ||
16918 isSHUFPMask(Mask, SVT) ||
16919 isSHUFPMask(Mask, SVT, /* Commuted */ true));
16924 //===----------------------------------------------------------------------===//
16925 // X86 Scheduler Hooks
16926 //===----------------------------------------------------------------------===//
16928 /// Utility function to emit xbegin specifying the start of an RTM region.
16929 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
16930 const TargetInstrInfo *TII) {
16931 DebugLoc DL = MI->getDebugLoc();
16933 const BasicBlock *BB = MBB->getBasicBlock();
16934 MachineFunction::iterator I = MBB;
16937 // For the v = xbegin(), we generate
16948 MachineBasicBlock *thisMBB = MBB;
16949 MachineFunction *MF = MBB->getParent();
16950 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
16951 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
16952 MF->insert(I, mainMBB);
16953 MF->insert(I, sinkMBB);
16955 // Transfer the remainder of BB and its successor edges to sinkMBB.
16956 sinkMBB->splice(sinkMBB->begin(), MBB,
16957 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
16958 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
16962 // # fallthrough to mainMBB
16963 // # abortion to sinkMBB
16964 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
16965 thisMBB->addSuccessor(mainMBB);
16966 thisMBB->addSuccessor(sinkMBB);
16970 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
16971 mainMBB->addSuccessor(sinkMBB);
16974 // EAX is live into the sinkMBB
16975 sinkMBB->addLiveIn(X86::EAX);
16976 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
16977 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
16980 MI->eraseFromParent();
16984 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
16985 // or XMM0_V32I8 in AVX all of this code can be replaced with that
16986 // in the .td file.
16987 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
16988 const TargetInstrInfo *TII) {
16990 switch (MI->getOpcode()) {
16991 default: llvm_unreachable("illegal opcode!");
16992 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
16993 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
16994 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
16995 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
16996 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
16997 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
16998 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
16999 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
17002 DebugLoc dl = MI->getDebugLoc();
17003 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
17005 unsigned NumArgs = MI->getNumOperands();
17006 for (unsigned i = 1; i < NumArgs; ++i) {
17007 MachineOperand &Op = MI->getOperand(i);
17008 if (!(Op.isReg() && Op.isImplicit()))
17009 MIB.addOperand(Op);
17011 if (MI->hasOneMemOperand())
17012 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
17014 BuildMI(*BB, MI, dl,
17015 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17016 .addReg(X86::XMM0);
17018 MI->eraseFromParent();
17022 // FIXME: Custom handling because TableGen doesn't support multiple implicit
17023 // defs in an instruction pattern
17024 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
17025 const TargetInstrInfo *TII) {
17027 switch (MI->getOpcode()) {
17028 default: llvm_unreachable("illegal opcode!");
17029 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
17030 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
17031 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
17032 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
17033 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
17034 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
17035 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
17036 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
17039 DebugLoc dl = MI->getDebugLoc();
17040 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
17042 unsigned NumArgs = MI->getNumOperands(); // remove the results
17043 for (unsigned i = 1; i < NumArgs; ++i) {
17044 MachineOperand &Op = MI->getOperand(i);
17045 if (!(Op.isReg() && Op.isImplicit()))
17046 MIB.addOperand(Op);
17048 if (MI->hasOneMemOperand())
17049 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
17051 BuildMI(*BB, MI, dl,
17052 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17055 MI->eraseFromParent();
17059 static MachineBasicBlock * EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
17060 const TargetInstrInfo *TII,
17061 const X86Subtarget* Subtarget) {
17062 DebugLoc dl = MI->getDebugLoc();
17064 // Address into RAX/EAX, other two args into ECX, EDX.
17065 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
17066 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
17067 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
17068 for (int i = 0; i < X86::AddrNumOperands; ++i)
17069 MIB.addOperand(MI->getOperand(i));
17071 unsigned ValOps = X86::AddrNumOperands;
17072 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
17073 .addReg(MI->getOperand(ValOps).getReg());
17074 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
17075 .addReg(MI->getOperand(ValOps+1).getReg());
17077 // The instruction doesn't actually take any operands though.
17078 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
17080 MI->eraseFromParent(); // The pseudo is gone now.
17084 MachineBasicBlock *
17085 X86TargetLowering::EmitVAARG64WithCustomInserter(
17087 MachineBasicBlock *MBB) const {
17088 // Emit va_arg instruction on X86-64.
17090 // Operands to this pseudo-instruction:
17091 // 0 ) Output : destination address (reg)
17092 // 1-5) Input : va_list address (addr, i64mem)
17093 // 6 ) ArgSize : Size (in bytes) of vararg type
17094 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
17095 // 8 ) Align : Alignment of type
17096 // 9 ) EFLAGS (implicit-def)
17098 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
17099 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
17101 unsigned DestReg = MI->getOperand(0).getReg();
17102 MachineOperand &Base = MI->getOperand(1);
17103 MachineOperand &Scale = MI->getOperand(2);
17104 MachineOperand &Index = MI->getOperand(3);
17105 MachineOperand &Disp = MI->getOperand(4);
17106 MachineOperand &Segment = MI->getOperand(5);
17107 unsigned ArgSize = MI->getOperand(6).getImm();
17108 unsigned ArgMode = MI->getOperand(7).getImm();
17109 unsigned Align = MI->getOperand(8).getImm();
17111 // Memory Reference
17112 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
17113 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
17114 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
17116 // Machine Information
17117 const TargetInstrInfo *TII = MBB->getParent()->getTarget().getInstrInfo();
17118 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
17119 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
17120 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
17121 DebugLoc DL = MI->getDebugLoc();
17123 // struct va_list {
17126 // i64 overflow_area (address)
17127 // i64 reg_save_area (address)
17129 // sizeof(va_list) = 24
17130 // alignment(va_list) = 8
17132 unsigned TotalNumIntRegs = 6;
17133 unsigned TotalNumXMMRegs = 8;
17134 bool UseGPOffset = (ArgMode == 1);
17135 bool UseFPOffset = (ArgMode == 2);
17136 unsigned MaxOffset = TotalNumIntRegs * 8 +
17137 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
17139 /* Align ArgSize to a multiple of 8 */
17140 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
17141 bool NeedsAlign = (Align > 8);
17143 MachineBasicBlock *thisMBB = MBB;
17144 MachineBasicBlock *overflowMBB;
17145 MachineBasicBlock *offsetMBB;
17146 MachineBasicBlock *endMBB;
17148 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
17149 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
17150 unsigned OffsetReg = 0;
17152 if (!UseGPOffset && !UseFPOffset) {
17153 // If we only pull from the overflow region, we don't create a branch.
17154 // We don't need to alter control flow.
17155 OffsetDestReg = 0; // unused
17156 OverflowDestReg = DestReg;
17158 offsetMBB = nullptr;
17159 overflowMBB = thisMBB;
17162 // First emit code to check if gp_offset (or fp_offset) is below the bound.
17163 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
17164 // If not, pull from overflow_area. (branch to overflowMBB)
17169 // offsetMBB overflowMBB
17174 // Registers for the PHI in endMBB
17175 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
17176 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
17178 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
17179 MachineFunction *MF = MBB->getParent();
17180 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17181 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17182 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17184 MachineFunction::iterator MBBIter = MBB;
17187 // Insert the new basic blocks
17188 MF->insert(MBBIter, offsetMBB);
17189 MF->insert(MBBIter, overflowMBB);
17190 MF->insert(MBBIter, endMBB);
17192 // Transfer the remainder of MBB and its successor edges to endMBB.
17193 endMBB->splice(endMBB->begin(), thisMBB,
17194 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
17195 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
17197 // Make offsetMBB and overflowMBB successors of thisMBB
17198 thisMBB->addSuccessor(offsetMBB);
17199 thisMBB->addSuccessor(overflowMBB);
17201 // endMBB is a successor of both offsetMBB and overflowMBB
17202 offsetMBB->addSuccessor(endMBB);
17203 overflowMBB->addSuccessor(endMBB);
17205 // Load the offset value into a register
17206 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
17207 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
17211 .addDisp(Disp, UseFPOffset ? 4 : 0)
17212 .addOperand(Segment)
17213 .setMemRefs(MMOBegin, MMOEnd);
17215 // Check if there is enough room left to pull this argument.
17216 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
17218 .addImm(MaxOffset + 8 - ArgSizeA8);
17220 // Branch to "overflowMBB" if offset >= max
17221 // Fall through to "offsetMBB" otherwise
17222 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
17223 .addMBB(overflowMBB);
17226 // In offsetMBB, emit code to use the reg_save_area.
17228 assert(OffsetReg != 0);
17230 // Read the reg_save_area address.
17231 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
17232 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
17237 .addOperand(Segment)
17238 .setMemRefs(MMOBegin, MMOEnd);
17240 // Zero-extend the offset
17241 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
17242 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
17245 .addImm(X86::sub_32bit);
17247 // Add the offset to the reg_save_area to get the final address.
17248 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
17249 .addReg(OffsetReg64)
17250 .addReg(RegSaveReg);
17252 // Compute the offset for the next argument
17253 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
17254 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
17256 .addImm(UseFPOffset ? 16 : 8);
17258 // Store it back into the va_list.
17259 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
17263 .addDisp(Disp, UseFPOffset ? 4 : 0)
17264 .addOperand(Segment)
17265 .addReg(NextOffsetReg)
17266 .setMemRefs(MMOBegin, MMOEnd);
17269 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
17274 // Emit code to use overflow area
17277 // Load the overflow_area address into a register.
17278 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
17279 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
17284 .addOperand(Segment)
17285 .setMemRefs(MMOBegin, MMOEnd);
17287 // If we need to align it, do so. Otherwise, just copy the address
17288 // to OverflowDestReg.
17290 // Align the overflow address
17291 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
17292 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
17294 // aligned_addr = (addr + (align-1)) & ~(align-1)
17295 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
17296 .addReg(OverflowAddrReg)
17299 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
17301 .addImm(~(uint64_t)(Align-1));
17303 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
17304 .addReg(OverflowAddrReg);
17307 // Compute the next overflow address after this argument.
17308 // (the overflow address should be kept 8-byte aligned)
17309 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
17310 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
17311 .addReg(OverflowDestReg)
17312 .addImm(ArgSizeA8);
17314 // Store the new overflow address.
17315 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
17320 .addOperand(Segment)
17321 .addReg(NextAddrReg)
17322 .setMemRefs(MMOBegin, MMOEnd);
17324 // If we branched, emit the PHI to the front of endMBB.
17326 BuildMI(*endMBB, endMBB->begin(), DL,
17327 TII->get(X86::PHI), DestReg)
17328 .addReg(OffsetDestReg).addMBB(offsetMBB)
17329 .addReg(OverflowDestReg).addMBB(overflowMBB);
17332 // Erase the pseudo instruction
17333 MI->eraseFromParent();
17338 MachineBasicBlock *
17339 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
17341 MachineBasicBlock *MBB) const {
17342 // Emit code to save XMM registers to the stack. The ABI says that the
17343 // number of registers to save is given in %al, so it's theoretically
17344 // possible to do an indirect jump trick to avoid saving all of them,
17345 // however this code takes a simpler approach and just executes all
17346 // of the stores if %al is non-zero. It's less code, and it's probably
17347 // easier on the hardware branch predictor, and stores aren't all that
17348 // expensive anyway.
17350 // Create the new basic blocks. One block contains all the XMM stores,
17351 // and one block is the final destination regardless of whether any
17352 // stores were performed.
17353 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
17354 MachineFunction *F = MBB->getParent();
17355 MachineFunction::iterator MBBIter = MBB;
17357 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
17358 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
17359 F->insert(MBBIter, XMMSaveMBB);
17360 F->insert(MBBIter, EndMBB);
17362 // Transfer the remainder of MBB and its successor edges to EndMBB.
17363 EndMBB->splice(EndMBB->begin(), MBB,
17364 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
17365 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
17367 // The original block will now fall through to the XMM save block.
17368 MBB->addSuccessor(XMMSaveMBB);
17369 // The XMMSaveMBB will fall through to the end block.
17370 XMMSaveMBB->addSuccessor(EndMBB);
17372 // Now add the instructions.
17373 const TargetInstrInfo *TII = MBB->getParent()->getTarget().getInstrInfo();
17374 DebugLoc DL = MI->getDebugLoc();
17376 unsigned CountReg = MI->getOperand(0).getReg();
17377 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
17378 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
17380 if (!Subtarget->isTargetWin64()) {
17381 // If %al is 0, branch around the XMM save block.
17382 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
17383 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
17384 MBB->addSuccessor(EndMBB);
17387 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
17388 // that was just emitted, but clearly shouldn't be "saved".
17389 assert((MI->getNumOperands() <= 3 ||
17390 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
17391 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
17392 && "Expected last argument to be EFLAGS");
17393 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
17394 // In the XMM save block, save all the XMM argument registers.
17395 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
17396 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
17397 MachineMemOperand *MMO =
17398 F->getMachineMemOperand(
17399 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
17400 MachineMemOperand::MOStore,
17401 /*Size=*/16, /*Align=*/16);
17402 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
17403 .addFrameIndex(RegSaveFrameIndex)
17404 .addImm(/*Scale=*/1)
17405 .addReg(/*IndexReg=*/0)
17406 .addImm(/*Disp=*/Offset)
17407 .addReg(/*Segment=*/0)
17408 .addReg(MI->getOperand(i).getReg())
17409 .addMemOperand(MMO);
17412 MI->eraseFromParent(); // The pseudo instruction is gone now.
17417 // The EFLAGS operand of SelectItr might be missing a kill marker
17418 // because there were multiple uses of EFLAGS, and ISel didn't know
17419 // which to mark. Figure out whether SelectItr should have had a
17420 // kill marker, and set it if it should. Returns the correct kill
17422 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
17423 MachineBasicBlock* BB,
17424 const TargetRegisterInfo* TRI) {
17425 // Scan forward through BB for a use/def of EFLAGS.
17426 MachineBasicBlock::iterator miI(std::next(SelectItr));
17427 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
17428 const MachineInstr& mi = *miI;
17429 if (mi.readsRegister(X86::EFLAGS))
17431 if (mi.definesRegister(X86::EFLAGS))
17432 break; // Should have kill-flag - update below.
17435 // If we hit the end of the block, check whether EFLAGS is live into a
17437 if (miI == BB->end()) {
17438 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
17439 sEnd = BB->succ_end();
17440 sItr != sEnd; ++sItr) {
17441 MachineBasicBlock* succ = *sItr;
17442 if (succ->isLiveIn(X86::EFLAGS))
17447 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
17448 // out. SelectMI should have a kill flag on EFLAGS.
17449 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
17453 MachineBasicBlock *
17454 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
17455 MachineBasicBlock *BB) const {
17456 const TargetInstrInfo *TII = BB->getParent()->getTarget().getInstrInfo();
17457 DebugLoc DL = MI->getDebugLoc();
17459 // To "insert" a SELECT_CC instruction, we actually have to insert the
17460 // diamond control-flow pattern. The incoming instruction knows the
17461 // destination vreg to set, the condition code register to branch on, the
17462 // true/false values to select between, and a branch opcode to use.
17463 const BasicBlock *LLVM_BB = BB->getBasicBlock();
17464 MachineFunction::iterator It = BB;
17470 // cmpTY ccX, r1, r2
17472 // fallthrough --> copy0MBB
17473 MachineBasicBlock *thisMBB = BB;
17474 MachineFunction *F = BB->getParent();
17475 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
17476 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
17477 F->insert(It, copy0MBB);
17478 F->insert(It, sinkMBB);
17480 // If the EFLAGS register isn't dead in the terminator, then claim that it's
17481 // live into the sink and copy blocks.
17482 const TargetRegisterInfo* TRI = BB->getParent()->getTarget().getRegisterInfo();
17483 if (!MI->killsRegister(X86::EFLAGS) &&
17484 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
17485 copy0MBB->addLiveIn(X86::EFLAGS);
17486 sinkMBB->addLiveIn(X86::EFLAGS);
17489 // Transfer the remainder of BB and its successor edges to sinkMBB.
17490 sinkMBB->splice(sinkMBB->begin(), BB,
17491 std::next(MachineBasicBlock::iterator(MI)), BB->end());
17492 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
17494 // Add the true and fallthrough blocks as its successors.
17495 BB->addSuccessor(copy0MBB);
17496 BB->addSuccessor(sinkMBB);
17498 // Create the conditional branch instruction.
17500 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
17501 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
17504 // %FalseValue = ...
17505 // # fallthrough to sinkMBB
17506 copy0MBB->addSuccessor(sinkMBB);
17509 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
17511 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
17512 TII->get(X86::PHI), MI->getOperand(0).getReg())
17513 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
17514 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
17516 MI->eraseFromParent(); // The pseudo instruction is gone now.
17520 MachineBasicBlock *
17521 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, MachineBasicBlock *BB,
17522 bool Is64Bit) const {
17523 MachineFunction *MF = BB->getParent();
17524 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
17525 DebugLoc DL = MI->getDebugLoc();
17526 const BasicBlock *LLVM_BB = BB->getBasicBlock();
17528 assert(MF->shouldSplitStack());
17530 unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
17531 unsigned TlsOffset = Is64Bit ? 0x70 : 0x30;
17534 // ... [Till the alloca]
17535 // If stacklet is not large enough, jump to mallocMBB
17538 // Allocate by subtracting from RSP
17539 // Jump to continueMBB
17542 // Allocate by call to runtime
17546 // [rest of original BB]
17549 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17550 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17551 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17553 MachineRegisterInfo &MRI = MF->getRegInfo();
17554 const TargetRegisterClass *AddrRegClass =
17555 getRegClassFor(Is64Bit ? MVT::i64:MVT::i32);
17557 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
17558 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
17559 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
17560 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
17561 sizeVReg = MI->getOperand(1).getReg(),
17562 physSPReg = Is64Bit ? X86::RSP : X86::ESP;
17564 MachineFunction::iterator MBBIter = BB;
17567 MF->insert(MBBIter, bumpMBB);
17568 MF->insert(MBBIter, mallocMBB);
17569 MF->insert(MBBIter, continueMBB);
17571 continueMBB->splice(continueMBB->begin(), BB,
17572 std::next(MachineBasicBlock::iterator(MI)), BB->end());
17573 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
17575 // Add code to the main basic block to check if the stack limit has been hit,
17576 // and if so, jump to mallocMBB otherwise to bumpMBB.
17577 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
17578 BuildMI(BB, DL, TII->get(Is64Bit ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
17579 .addReg(tmpSPVReg).addReg(sizeVReg);
17580 BuildMI(BB, DL, TII->get(Is64Bit ? X86::CMP64mr:X86::CMP32mr))
17581 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
17582 .addReg(SPLimitVReg);
17583 BuildMI(BB, DL, TII->get(X86::JG_4)).addMBB(mallocMBB);
17585 // bumpMBB simply decreases the stack pointer, since we know the current
17586 // stacklet has enough space.
17587 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
17588 .addReg(SPLimitVReg);
17589 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
17590 .addReg(SPLimitVReg);
17591 BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
17593 // Calls into a routine in libgcc to allocate more space from the heap.
17594 const uint32_t *RegMask =
17595 MF->getTarget().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
17597 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
17599 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
17600 .addExternalSymbol("__morestack_allocate_stack_space")
17601 .addRegMask(RegMask)
17602 .addReg(X86::RDI, RegState::Implicit)
17603 .addReg(X86::RAX, RegState::ImplicitDefine);
17605 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
17607 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
17608 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
17609 .addExternalSymbol("__morestack_allocate_stack_space")
17610 .addRegMask(RegMask)
17611 .addReg(X86::EAX, RegState::ImplicitDefine);
17615 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
17618 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
17619 .addReg(Is64Bit ? X86::RAX : X86::EAX);
17620 BuildMI(mallocMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
17622 // Set up the CFG correctly.
17623 BB->addSuccessor(bumpMBB);
17624 BB->addSuccessor(mallocMBB);
17625 mallocMBB->addSuccessor(continueMBB);
17626 bumpMBB->addSuccessor(continueMBB);
17628 // Take care of the PHI nodes.
17629 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
17630 MI->getOperand(0).getReg())
17631 .addReg(mallocPtrVReg).addMBB(mallocMBB)
17632 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
17634 // Delete the original pseudo instruction.
17635 MI->eraseFromParent();
17638 return continueMBB;
17641 MachineBasicBlock *
17642 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
17643 MachineBasicBlock *BB) const {
17644 const TargetInstrInfo *TII = BB->getParent()->getTarget().getInstrInfo();
17645 DebugLoc DL = MI->getDebugLoc();
17647 assert(!Subtarget->isTargetMacho());
17649 // The lowering is pretty easy: we're just emitting the call to _alloca. The
17650 // non-trivial part is impdef of ESP.
17652 if (Subtarget->isTargetWin64()) {
17653 if (Subtarget->isTargetCygMing()) {
17654 // ___chkstk(Mingw64):
17655 // Clobbers R10, R11, RAX and EFLAGS.
17657 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
17658 .addExternalSymbol("___chkstk")
17659 .addReg(X86::RAX, RegState::Implicit)
17660 .addReg(X86::RSP, RegState::Implicit)
17661 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
17662 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
17663 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
17665 // __chkstk(MSVCRT): does not update stack pointer.
17666 // Clobbers R10, R11 and EFLAGS.
17667 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
17668 .addExternalSymbol("__chkstk")
17669 .addReg(X86::RAX, RegState::Implicit)
17670 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
17671 // RAX has the offset to be subtracted from RSP.
17672 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
17677 const char *StackProbeSymbol =
17678 Subtarget->isTargetKnownWindowsMSVC() ? "_chkstk" : "_alloca";
17680 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
17681 .addExternalSymbol(StackProbeSymbol)
17682 .addReg(X86::EAX, RegState::Implicit)
17683 .addReg(X86::ESP, RegState::Implicit)
17684 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
17685 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
17686 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
17689 MI->eraseFromParent(); // The pseudo instruction is gone now.
17693 MachineBasicBlock *
17694 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
17695 MachineBasicBlock *BB) const {
17696 // This is pretty easy. We're taking the value that we received from
17697 // our load from the relocation, sticking it in either RDI (x86-64)
17698 // or EAX and doing an indirect call. The return value will then
17699 // be in the normal return register.
17700 MachineFunction *F = BB->getParent();
17701 const X86InstrInfo *TII
17702 = static_cast<const X86InstrInfo*>(F->getTarget().getInstrInfo());
17703 DebugLoc DL = MI->getDebugLoc();
17705 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
17706 assert(MI->getOperand(3).isGlobal() && "This should be a global");
17708 // Get a register mask for the lowered call.
17709 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
17710 // proper register mask.
17711 const uint32_t *RegMask =
17712 F->getTarget().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
17713 if (Subtarget->is64Bit()) {
17714 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
17715 TII->get(X86::MOV64rm), X86::RDI)
17717 .addImm(0).addReg(0)
17718 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
17719 MI->getOperand(3).getTargetFlags())
17721 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
17722 addDirectMem(MIB, X86::RDI);
17723 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
17724 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
17725 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
17726 TII->get(X86::MOV32rm), X86::EAX)
17728 .addImm(0).addReg(0)
17729 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
17730 MI->getOperand(3).getTargetFlags())
17732 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
17733 addDirectMem(MIB, X86::EAX);
17734 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
17736 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
17737 TII->get(X86::MOV32rm), X86::EAX)
17738 .addReg(TII->getGlobalBaseReg(F))
17739 .addImm(0).addReg(0)
17740 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
17741 MI->getOperand(3).getTargetFlags())
17743 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
17744 addDirectMem(MIB, X86::EAX);
17745 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
17748 MI->eraseFromParent(); // The pseudo instruction is gone now.
17752 MachineBasicBlock *
17753 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
17754 MachineBasicBlock *MBB) const {
17755 DebugLoc DL = MI->getDebugLoc();
17756 MachineFunction *MF = MBB->getParent();
17757 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
17758 MachineRegisterInfo &MRI = MF->getRegInfo();
17760 const BasicBlock *BB = MBB->getBasicBlock();
17761 MachineFunction::iterator I = MBB;
17764 // Memory Reference
17765 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
17766 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
17769 unsigned MemOpndSlot = 0;
17771 unsigned CurOp = 0;
17773 DstReg = MI->getOperand(CurOp++).getReg();
17774 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
17775 assert(RC->hasType(MVT::i32) && "Invalid destination!");
17776 unsigned mainDstReg = MRI.createVirtualRegister(RC);
17777 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
17779 MemOpndSlot = CurOp;
17781 MVT PVT = getPointerTy();
17782 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
17783 "Invalid Pointer Size!");
17785 // For v = setjmp(buf), we generate
17788 // buf[LabelOffset] = restoreMBB
17789 // SjLjSetup restoreMBB
17795 // v = phi(main, restore)
17800 MachineBasicBlock *thisMBB = MBB;
17801 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
17802 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
17803 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
17804 MF->insert(I, mainMBB);
17805 MF->insert(I, sinkMBB);
17806 MF->push_back(restoreMBB);
17808 MachineInstrBuilder MIB;
17810 // Transfer the remainder of BB and its successor edges to sinkMBB.
17811 sinkMBB->splice(sinkMBB->begin(), MBB,
17812 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
17813 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
17816 unsigned PtrStoreOpc = 0;
17817 unsigned LabelReg = 0;
17818 const int64_t LabelOffset = 1 * PVT.getStoreSize();
17819 Reloc::Model RM = MF->getTarget().getRelocationModel();
17820 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
17821 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
17823 // Prepare IP either in reg or imm.
17824 if (!UseImmLabel) {
17825 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
17826 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
17827 LabelReg = MRI.createVirtualRegister(PtrRC);
17828 if (Subtarget->is64Bit()) {
17829 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
17833 .addMBB(restoreMBB)
17836 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
17837 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
17838 .addReg(XII->getGlobalBaseReg(MF))
17841 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
17845 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
17847 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
17848 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
17849 if (i == X86::AddrDisp)
17850 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
17852 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
17855 MIB.addReg(LabelReg);
17857 MIB.addMBB(restoreMBB);
17858 MIB.setMemRefs(MMOBegin, MMOEnd);
17860 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
17861 .addMBB(restoreMBB);
17863 const X86RegisterInfo *RegInfo =
17864 static_cast<const X86RegisterInfo*>(MF->getTarget().getRegisterInfo());
17865 MIB.addRegMask(RegInfo->getNoPreservedMask());
17866 thisMBB->addSuccessor(mainMBB);
17867 thisMBB->addSuccessor(restoreMBB);
17871 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
17872 mainMBB->addSuccessor(sinkMBB);
17875 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
17876 TII->get(X86::PHI), DstReg)
17877 .addReg(mainDstReg).addMBB(mainMBB)
17878 .addReg(restoreDstReg).addMBB(restoreMBB);
17881 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
17882 BuildMI(restoreMBB, DL, TII->get(X86::JMP_4)).addMBB(sinkMBB);
17883 restoreMBB->addSuccessor(sinkMBB);
17885 MI->eraseFromParent();
17889 MachineBasicBlock *
17890 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
17891 MachineBasicBlock *MBB) const {
17892 DebugLoc DL = MI->getDebugLoc();
17893 MachineFunction *MF = MBB->getParent();
17894 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
17895 MachineRegisterInfo &MRI = MF->getRegInfo();
17897 // Memory Reference
17898 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
17899 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
17901 MVT PVT = getPointerTy();
17902 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
17903 "Invalid Pointer Size!");
17905 const TargetRegisterClass *RC =
17906 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
17907 unsigned Tmp = MRI.createVirtualRegister(RC);
17908 // Since FP is only updated here but NOT referenced, it's treated as GPR.
17909 const X86RegisterInfo *RegInfo =
17910 static_cast<const X86RegisterInfo*>(MF->getTarget().getRegisterInfo());
17911 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
17912 unsigned SP = RegInfo->getStackRegister();
17914 MachineInstrBuilder MIB;
17916 const int64_t LabelOffset = 1 * PVT.getStoreSize();
17917 const int64_t SPOffset = 2 * PVT.getStoreSize();
17919 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
17920 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
17923 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
17924 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
17925 MIB.addOperand(MI->getOperand(i));
17926 MIB.setMemRefs(MMOBegin, MMOEnd);
17928 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
17929 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
17930 if (i == X86::AddrDisp)
17931 MIB.addDisp(MI->getOperand(i), LabelOffset);
17933 MIB.addOperand(MI->getOperand(i));
17935 MIB.setMemRefs(MMOBegin, MMOEnd);
17937 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
17938 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
17939 if (i == X86::AddrDisp)
17940 MIB.addDisp(MI->getOperand(i), SPOffset);
17942 MIB.addOperand(MI->getOperand(i));
17944 MIB.setMemRefs(MMOBegin, MMOEnd);
17946 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
17948 MI->eraseFromParent();
17952 // Replace 213-type (isel default) FMA3 instructions with 231-type for
17953 // accumulator loops. Writing back to the accumulator allows the coalescer
17954 // to remove extra copies in the loop.
17955 MachineBasicBlock *
17956 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
17957 MachineBasicBlock *MBB) const {
17958 MachineOperand &AddendOp = MI->getOperand(3);
17960 // Bail out early if the addend isn't a register - we can't switch these.
17961 if (!AddendOp.isReg())
17964 MachineFunction &MF = *MBB->getParent();
17965 MachineRegisterInfo &MRI = MF.getRegInfo();
17967 // Check whether the addend is defined by a PHI:
17968 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
17969 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
17970 if (!AddendDef.isPHI())
17973 // Look for the following pattern:
17975 // %addend = phi [%entry, 0], [%loop, %result]
17977 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
17981 // %addend = phi [%entry, 0], [%loop, %result]
17983 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
17985 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
17986 assert(AddendDef.getOperand(i).isReg());
17987 MachineOperand PHISrcOp = AddendDef.getOperand(i);
17988 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
17989 if (&PHISrcInst == MI) {
17990 // Found a matching instruction.
17991 unsigned NewFMAOpc = 0;
17992 switch (MI->getOpcode()) {
17993 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
17994 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
17995 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
17996 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
17997 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
17998 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
17999 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
18000 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
18001 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
18002 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
18003 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
18004 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
18005 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
18006 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
18007 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
18008 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
18009 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
18010 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
18011 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
18012 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
18013 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
18014 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
18015 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
18016 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
18017 default: llvm_unreachable("Unrecognized FMA variant.");
18020 const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
18021 MachineInstrBuilder MIB =
18022 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
18023 .addOperand(MI->getOperand(0))
18024 .addOperand(MI->getOperand(3))
18025 .addOperand(MI->getOperand(2))
18026 .addOperand(MI->getOperand(1));
18027 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
18028 MI->eraseFromParent();
18035 MachineBasicBlock *
18036 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
18037 MachineBasicBlock *BB) const {
18038 switch (MI->getOpcode()) {
18039 default: llvm_unreachable("Unexpected instr type to insert");
18040 case X86::TAILJMPd64:
18041 case X86::TAILJMPr64:
18042 case X86::TAILJMPm64:
18043 llvm_unreachable("TAILJMP64 would not be touched here.");
18044 case X86::TCRETURNdi64:
18045 case X86::TCRETURNri64:
18046 case X86::TCRETURNmi64:
18048 case X86::WIN_ALLOCA:
18049 return EmitLoweredWinAlloca(MI, BB);
18050 case X86::SEG_ALLOCA_32:
18051 return EmitLoweredSegAlloca(MI, BB, false);
18052 case X86::SEG_ALLOCA_64:
18053 return EmitLoweredSegAlloca(MI, BB, true);
18054 case X86::TLSCall_32:
18055 case X86::TLSCall_64:
18056 return EmitLoweredTLSCall(MI, BB);
18057 case X86::CMOV_GR8:
18058 case X86::CMOV_FR32:
18059 case X86::CMOV_FR64:
18060 case X86::CMOV_V4F32:
18061 case X86::CMOV_V2F64:
18062 case X86::CMOV_V2I64:
18063 case X86::CMOV_V8F32:
18064 case X86::CMOV_V4F64:
18065 case X86::CMOV_V4I64:
18066 case X86::CMOV_V16F32:
18067 case X86::CMOV_V8F64:
18068 case X86::CMOV_V8I64:
18069 case X86::CMOV_GR16:
18070 case X86::CMOV_GR32:
18071 case X86::CMOV_RFP32:
18072 case X86::CMOV_RFP64:
18073 case X86::CMOV_RFP80:
18074 return EmitLoweredSelect(MI, BB);
18076 case X86::FP32_TO_INT16_IN_MEM:
18077 case X86::FP32_TO_INT32_IN_MEM:
18078 case X86::FP32_TO_INT64_IN_MEM:
18079 case X86::FP64_TO_INT16_IN_MEM:
18080 case X86::FP64_TO_INT32_IN_MEM:
18081 case X86::FP64_TO_INT64_IN_MEM:
18082 case X86::FP80_TO_INT16_IN_MEM:
18083 case X86::FP80_TO_INT32_IN_MEM:
18084 case X86::FP80_TO_INT64_IN_MEM: {
18085 MachineFunction *F = BB->getParent();
18086 const TargetInstrInfo *TII = F->getTarget().getInstrInfo();
18087 DebugLoc DL = MI->getDebugLoc();
18089 // Change the floating point control register to use "round towards zero"
18090 // mode when truncating to an integer value.
18091 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
18092 addFrameReference(BuildMI(*BB, MI, DL,
18093 TII->get(X86::FNSTCW16m)), CWFrameIdx);
18095 // Load the old value of the high byte of the control word...
18097 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
18098 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
18101 // Set the high part to be round to zero...
18102 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
18105 // Reload the modified control word now...
18106 addFrameReference(BuildMI(*BB, MI, DL,
18107 TII->get(X86::FLDCW16m)), CWFrameIdx);
18109 // Restore the memory image of control word to original value
18110 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
18113 // Get the X86 opcode to use.
18115 switch (MI->getOpcode()) {
18116 default: llvm_unreachable("illegal opcode!");
18117 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
18118 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
18119 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
18120 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
18121 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
18122 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
18123 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
18124 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
18125 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
18129 MachineOperand &Op = MI->getOperand(0);
18131 AM.BaseType = X86AddressMode::RegBase;
18132 AM.Base.Reg = Op.getReg();
18134 AM.BaseType = X86AddressMode::FrameIndexBase;
18135 AM.Base.FrameIndex = Op.getIndex();
18137 Op = MI->getOperand(1);
18139 AM.Scale = Op.getImm();
18140 Op = MI->getOperand(2);
18142 AM.IndexReg = Op.getImm();
18143 Op = MI->getOperand(3);
18144 if (Op.isGlobal()) {
18145 AM.GV = Op.getGlobal();
18147 AM.Disp = Op.getImm();
18149 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
18150 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
18152 // Reload the original control word now.
18153 addFrameReference(BuildMI(*BB, MI, DL,
18154 TII->get(X86::FLDCW16m)), CWFrameIdx);
18156 MI->eraseFromParent(); // The pseudo instruction is gone now.
18159 // String/text processing lowering.
18160 case X86::PCMPISTRM128REG:
18161 case X86::VPCMPISTRM128REG:
18162 case X86::PCMPISTRM128MEM:
18163 case X86::VPCMPISTRM128MEM:
18164 case X86::PCMPESTRM128REG:
18165 case X86::VPCMPESTRM128REG:
18166 case X86::PCMPESTRM128MEM:
18167 case X86::VPCMPESTRM128MEM:
18168 assert(Subtarget->hasSSE42() &&
18169 "Target must have SSE4.2 or AVX features enabled");
18170 return EmitPCMPSTRM(MI, BB, BB->getParent()->getTarget().getInstrInfo());
18172 // String/text processing lowering.
18173 case X86::PCMPISTRIREG:
18174 case X86::VPCMPISTRIREG:
18175 case X86::PCMPISTRIMEM:
18176 case X86::VPCMPISTRIMEM:
18177 case X86::PCMPESTRIREG:
18178 case X86::VPCMPESTRIREG:
18179 case X86::PCMPESTRIMEM:
18180 case X86::VPCMPESTRIMEM:
18181 assert(Subtarget->hasSSE42() &&
18182 "Target must have SSE4.2 or AVX features enabled");
18183 return EmitPCMPSTRI(MI, BB, BB->getParent()->getTarget().getInstrInfo());
18185 // Thread synchronization.
18187 return EmitMonitor(MI, BB, BB->getParent()->getTarget().getInstrInfo(), Subtarget);
18191 return EmitXBegin(MI, BB, BB->getParent()->getTarget().getInstrInfo());
18193 case X86::VASTART_SAVE_XMM_REGS:
18194 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
18196 case X86::VAARG_64:
18197 return EmitVAARG64WithCustomInserter(MI, BB);
18199 case X86::EH_SjLj_SetJmp32:
18200 case X86::EH_SjLj_SetJmp64:
18201 return emitEHSjLjSetJmp(MI, BB);
18203 case X86::EH_SjLj_LongJmp32:
18204 case X86::EH_SjLj_LongJmp64:
18205 return emitEHSjLjLongJmp(MI, BB);
18207 case TargetOpcode::STACKMAP:
18208 case TargetOpcode::PATCHPOINT:
18209 return emitPatchPoint(MI, BB);
18211 case X86::VFMADDPDr213r:
18212 case X86::VFMADDPSr213r:
18213 case X86::VFMADDSDr213r:
18214 case X86::VFMADDSSr213r:
18215 case X86::VFMSUBPDr213r:
18216 case X86::VFMSUBPSr213r:
18217 case X86::VFMSUBSDr213r:
18218 case X86::VFMSUBSSr213r:
18219 case X86::VFNMADDPDr213r:
18220 case X86::VFNMADDPSr213r:
18221 case X86::VFNMADDSDr213r:
18222 case X86::VFNMADDSSr213r:
18223 case X86::VFNMSUBPDr213r:
18224 case X86::VFNMSUBPSr213r:
18225 case X86::VFNMSUBSDr213r:
18226 case X86::VFNMSUBSSr213r:
18227 case X86::VFMADDPDr213rY:
18228 case X86::VFMADDPSr213rY:
18229 case X86::VFMSUBPDr213rY:
18230 case X86::VFMSUBPSr213rY:
18231 case X86::VFNMADDPDr213rY:
18232 case X86::VFNMADDPSr213rY:
18233 case X86::VFNMSUBPDr213rY:
18234 case X86::VFNMSUBPSr213rY:
18235 return emitFMA3Instr(MI, BB);
18239 //===----------------------------------------------------------------------===//
18240 // X86 Optimization Hooks
18241 //===----------------------------------------------------------------------===//
18243 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
18246 const SelectionDAG &DAG,
18247 unsigned Depth) const {
18248 unsigned BitWidth = KnownZero.getBitWidth();
18249 unsigned Opc = Op.getOpcode();
18250 assert((Opc >= ISD::BUILTIN_OP_END ||
18251 Opc == ISD::INTRINSIC_WO_CHAIN ||
18252 Opc == ISD::INTRINSIC_W_CHAIN ||
18253 Opc == ISD::INTRINSIC_VOID) &&
18254 "Should use MaskedValueIsZero if you don't know whether Op"
18255 " is a target node!");
18257 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
18271 // These nodes' second result is a boolean.
18272 if (Op.getResNo() == 0)
18275 case X86ISD::SETCC:
18276 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
18278 case ISD::INTRINSIC_WO_CHAIN: {
18279 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
18280 unsigned NumLoBits = 0;
18283 case Intrinsic::x86_sse_movmsk_ps:
18284 case Intrinsic::x86_avx_movmsk_ps_256:
18285 case Intrinsic::x86_sse2_movmsk_pd:
18286 case Intrinsic::x86_avx_movmsk_pd_256:
18287 case Intrinsic::x86_mmx_pmovmskb:
18288 case Intrinsic::x86_sse2_pmovmskb_128:
18289 case Intrinsic::x86_avx2_pmovmskb: {
18290 // High bits of movmskp{s|d}, pmovmskb are known zero.
18292 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
18293 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
18294 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
18295 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
18296 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
18297 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
18298 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
18299 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
18301 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
18310 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
18312 const SelectionDAG &,
18313 unsigned Depth) const {
18314 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
18315 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
18316 return Op.getValueType().getScalarType().getSizeInBits();
18322 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
18323 /// node is a GlobalAddress + offset.
18324 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
18325 const GlobalValue* &GA,
18326 int64_t &Offset) const {
18327 if (N->getOpcode() == X86ISD::Wrapper) {
18328 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
18329 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
18330 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
18334 return TargetLowering::isGAPlusOffset(N, GA, Offset);
18337 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
18338 /// same as extracting the high 128-bit part of 256-bit vector and then
18339 /// inserting the result into the low part of a new 256-bit vector
18340 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
18341 EVT VT = SVOp->getValueType(0);
18342 unsigned NumElems = VT.getVectorNumElements();
18344 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
18345 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
18346 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
18347 SVOp->getMaskElt(j) >= 0)
18353 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
18354 /// same as extracting the low 128-bit part of 256-bit vector and then
18355 /// inserting the result into the high part of a new 256-bit vector
18356 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
18357 EVT VT = SVOp->getValueType(0);
18358 unsigned NumElems = VT.getVectorNumElements();
18360 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
18361 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
18362 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
18363 SVOp->getMaskElt(j) >= 0)
18369 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
18370 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
18371 TargetLowering::DAGCombinerInfo &DCI,
18372 const X86Subtarget* Subtarget) {
18374 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
18375 SDValue V1 = SVOp->getOperand(0);
18376 SDValue V2 = SVOp->getOperand(1);
18377 EVT VT = SVOp->getValueType(0);
18378 unsigned NumElems = VT.getVectorNumElements();
18380 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
18381 V2.getOpcode() == ISD::CONCAT_VECTORS) {
18385 // V UNDEF BUILD_VECTOR UNDEF
18387 // CONCAT_VECTOR CONCAT_VECTOR
18390 // RESULT: V + zero extended
18392 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
18393 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
18394 V1.getOperand(1).getOpcode() != ISD::UNDEF)
18397 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
18400 // To match the shuffle mask, the first half of the mask should
18401 // be exactly the first vector, and all the rest a splat with the
18402 // first element of the second one.
18403 for (unsigned i = 0; i != NumElems/2; ++i)
18404 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
18405 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
18408 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
18409 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
18410 if (Ld->hasNUsesOfValue(1, 0)) {
18411 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
18412 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
18414 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
18416 Ld->getPointerInfo(),
18417 Ld->getAlignment(),
18418 false/*isVolatile*/, true/*ReadMem*/,
18419 false/*WriteMem*/);
18421 // Make sure the newly-created LOAD is in the same position as Ld in
18422 // terms of dependency. We create a TokenFactor for Ld and ResNode,
18423 // and update uses of Ld's output chain to use the TokenFactor.
18424 if (Ld->hasAnyUseOfValue(1)) {
18425 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
18426 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
18427 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
18428 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
18429 SDValue(ResNode.getNode(), 1));
18432 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
18436 // Emit a zeroed vector and insert the desired subvector on its
18438 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18439 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
18440 return DCI.CombineTo(N, InsV);
18443 //===--------------------------------------------------------------------===//
18444 // Combine some shuffles into subvector extracts and inserts:
18447 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
18448 if (isShuffleHigh128VectorInsertLow(SVOp)) {
18449 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
18450 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
18451 return DCI.CombineTo(N, InsV);
18454 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
18455 if (isShuffleLow128VectorInsertHigh(SVOp)) {
18456 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
18457 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
18458 return DCI.CombineTo(N, InsV);
18464 /// \brief Get the PSHUF-style mask from PSHUF node.
18466 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
18467 /// PSHUF-style masks that can be reused with such instructions.
18468 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
18469 SmallVector<int, 4> Mask;
18471 bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary);
18475 switch (N.getOpcode()) {
18476 case X86ISD::PSHUFD:
18478 case X86ISD::PSHUFLW:
18481 case X86ISD::PSHUFHW:
18482 Mask.erase(Mask.begin(), Mask.begin() + 4);
18483 for (int &M : Mask)
18487 llvm_unreachable("No valid shuffle instruction found!");
18491 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
18493 /// We walk up the chain and look for a combinable shuffle, skipping over
18494 /// shuffles that we could hoist this shuffle's transformation past without
18495 /// altering anything.
18496 static bool combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
18498 TargetLowering::DAGCombinerInfo &DCI) {
18499 assert(N.getOpcode() == X86ISD::PSHUFD &&
18500 "Called with something other than an x86 128-bit half shuffle!");
18503 // Walk up a single-use chain looking for a combinable shuffle.
18504 SDValue V = N.getOperand(0);
18505 for (; V.hasOneUse(); V = V.getOperand(0)) {
18506 switch (V.getOpcode()) {
18508 return false; // Nothing combined!
18511 // Skip bitcasts as we always know the type for the target specific
18515 case X86ISD::PSHUFD:
18516 // Found another dword shuffle.
18519 case X86ISD::PSHUFLW:
18520 // Check that the low words (being shuffled) are the identity in the
18521 // dword shuffle, and the high words are self-contained.
18522 if (Mask[0] != 0 || Mask[1] != 1 ||
18523 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
18528 case X86ISD::PSHUFHW:
18529 // Check that the high words (being shuffled) are the identity in the
18530 // dword shuffle, and the low words are self-contained.
18531 if (Mask[2] != 2 || Mask[3] != 3 ||
18532 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
18537 case X86ISD::UNPCKL:
18538 case X86ISD::UNPCKH:
18539 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
18540 // shuffle into a preceding word shuffle.
18541 if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
18544 // Search for a half-shuffle which we can combine with.
18545 unsigned CombineOp =
18546 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
18547 if (V.getOperand(0) != V.getOperand(1) ||
18548 !V->isOnlyUserOf(V.getOperand(0).getNode()))
18550 V = V.getOperand(0);
18552 switch (V.getOpcode()) {
18554 return false; // Nothing to combine.
18556 case X86ISD::PSHUFLW:
18557 case X86ISD::PSHUFHW:
18558 if (V.getOpcode() == CombineOp)
18563 V = V.getOperand(0);
18567 } while (V.hasOneUse());
18570 // Break out of the loop if we break out of the switch.
18574 if (!V.hasOneUse())
18575 // We fell out of the loop without finding a viable combining instruction.
18578 // Record the old value to use in RAUW-ing.
18581 // Merge this node's mask and our incoming mask.
18582 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
18583 for (int &M : Mask)
18585 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
18586 getV4X86ShuffleImm8ForMask(Mask, DAG));
18588 // It is possible that one of the combinable shuffles was completely absorbed
18589 // by the other, just replace it and revisit all users in that case.
18590 if (Old.getNode() == V.getNode()) {
18591 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo=*/true);
18595 // Replace N with its operand as we're going to combine that shuffle away.
18596 DAG.ReplaceAllUsesWith(N, N.getOperand(0));
18598 // Replace the combinable shuffle with the combined one, updating all users
18599 // so that we re-evaluate the chain here.
18600 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
18604 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
18606 /// We walk up the chain, skipping shuffles of the other half and looking
18607 /// through shuffles which switch halves trying to find a shuffle of the same
18608 /// pair of dwords.
18609 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
18611 TargetLowering::DAGCombinerInfo &DCI) {
18613 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
18614 "Called with something other than an x86 128-bit half shuffle!");
18616 unsigned CombineOpcode = N.getOpcode();
18618 // Walk up a single-use chain looking for a combinable shuffle.
18619 SDValue V = N.getOperand(0);
18620 for (; V.hasOneUse(); V = V.getOperand(0)) {
18621 switch (V.getOpcode()) {
18623 return false; // Nothing combined!
18626 // Skip bitcasts as we always know the type for the target specific
18630 case X86ISD::PSHUFLW:
18631 case X86ISD::PSHUFHW:
18632 if (V.getOpcode() == CombineOpcode)
18635 // Other-half shuffles are no-ops.
18638 case X86ISD::PSHUFD: {
18639 // We can only handle pshufd if the half we are combining either stays in
18640 // its half, or switches to the other half. Bail if one of these isn't
18642 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
18643 int DOffset = CombineOpcode == X86ISD::PSHUFLW ? 0 : 2;
18644 if (!((VMask[DOffset + 0] < 2 && VMask[DOffset + 1] < 2) ||
18645 (VMask[DOffset + 0] >= 2 && VMask[DOffset + 1] >= 2)))
18648 // Map the mask through the pshufd and keep walking up the chain.
18649 for (int i = 0; i < 4; ++i)
18650 Mask[i] = 2 * (VMask[DOffset + Mask[i] / 2] % 2) + Mask[i] % 2;
18652 // Switch halves if the pshufd does.
18654 VMask[DOffset + Mask[0] / 2] < 2 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
18658 // Break out of the loop if we break out of the switch.
18662 if (!V.hasOneUse())
18663 // We fell out of the loop without finding a viable combining instruction.
18666 // Record the old value to use in RAUW-ing.
18669 // Merge this node's mask and our incoming mask (adjusted to account for all
18670 // the pshufd instructions encountered).
18671 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
18672 for (int &M : Mask)
18674 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
18675 getV4X86ShuffleImm8ForMask(Mask, DAG));
18677 // Replace N with its operand as we're going to combine that shuffle away.
18678 DAG.ReplaceAllUsesWith(N, N.getOperand(0));
18680 // Replace the combinable shuffle with the combined one, updating all users
18681 // so that we re-evaluate the chain here.
18682 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
18686 /// \brief Try to combine x86 target specific shuffles.
18687 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
18688 TargetLowering::DAGCombinerInfo &DCI,
18689 const X86Subtarget *Subtarget) {
18691 MVT VT = N.getSimpleValueType();
18692 SmallVector<int, 4> Mask;
18694 switch (N.getOpcode()) {
18695 case X86ISD::PSHUFD:
18696 case X86ISD::PSHUFLW:
18697 case X86ISD::PSHUFHW:
18698 Mask = getPSHUFShuffleMask(N);
18699 assert(Mask.size() == 4);
18705 // Nuke no-op shuffles that show up after combining.
18706 if (isNoopShuffleMask(Mask))
18707 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
18709 // Look for simplifications involving one or two shuffle instructions.
18710 SDValue V = N.getOperand(0);
18711 switch (N.getOpcode()) {
18714 case X86ISD::PSHUFLW:
18715 case X86ISD::PSHUFHW:
18716 assert(VT == MVT::v8i16);
18719 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
18720 return SDValue(); // We combined away this shuffle, so we're done.
18722 // See if this reduces to a PSHUFD which is no more expensive and can
18723 // combine with more operations.
18724 if (Mask[0] % 2 == 0 && Mask[2] % 2 == 0 &&
18725 areAdjacentMasksSequential(Mask)) {
18726 int DMask[] = {-1, -1, -1, -1};
18727 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
18728 DMask[DOffset + 0] = DOffset + Mask[0] / 2;
18729 DMask[DOffset + 1] = DOffset + Mask[2] / 2;
18730 V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V);
18731 DCI.AddToWorklist(V.getNode());
18732 V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V,
18733 getV4X86ShuffleImm8ForMask(DMask, DAG));
18734 DCI.AddToWorklist(V.getNode());
18735 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
18738 // Look for shuffle patterns which can be implemented as a single unpack.
18739 // FIXME: This doesn't handle the location of the PSHUFD generically, and
18740 // only works when we have a PSHUFD followed by two half-shuffles.
18741 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
18742 (V.getOpcode() == X86ISD::PSHUFLW ||
18743 V.getOpcode() == X86ISD::PSHUFHW) &&
18744 V.getOpcode() != N.getOpcode() &&
18746 SDValue D = V.getOperand(0);
18747 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
18748 D = D.getOperand(0);
18749 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
18750 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
18751 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
18752 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
18753 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
18755 for (int i = 0; i < 4; ++i) {
18756 WordMask[i + NOffset] = Mask[i] + NOffset;
18757 WordMask[i + VOffset] = VMask[i] + VOffset;
18759 // Map the word mask through the DWord mask.
18761 for (int i = 0; i < 8; ++i)
18762 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
18763 const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3};
18764 const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7};
18765 if (std::equal(std::begin(MappedMask), std::end(MappedMask),
18766 std::begin(UnpackLoMask)) ||
18767 std::equal(std::begin(MappedMask), std::end(MappedMask),
18768 std::begin(UnpackHiMask))) {
18769 // We can replace all three shuffles with an unpack.
18770 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
18771 DCI.AddToWorklist(V.getNode());
18772 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
18774 DL, MVT::v8i16, V, V);
18781 case X86ISD::PSHUFD:
18782 if (combineRedundantDWordShuffle(N, Mask, DAG, DCI))
18783 return SDValue(); // We combined away this shuffle.
18791 /// PerformShuffleCombine - Performs several different shuffle combines.
18792 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
18793 TargetLowering::DAGCombinerInfo &DCI,
18794 const X86Subtarget *Subtarget) {
18796 SDValue N0 = N->getOperand(0);
18797 SDValue N1 = N->getOperand(1);
18798 EVT VT = N->getValueType(0);
18800 // Canonicalize shuffles that perform 'addsub' on packed float vectors
18801 // according to the rule:
18802 // (shuffle (FADD A, B), (FSUB A, B), Mask) ->
18803 // (shuffle (FSUB A, -B), (FADD A, -B), Mask)
18805 // Where 'Mask' is:
18806 // <0,5,2,7> -- for v4f32 and v4f64 shuffles;
18807 // <0,3> -- for v2f64 shuffles;
18808 // <0,9,2,11,4,13,6,15> -- for v8f32 shuffles.
18810 // This helps pattern-matching more SSE3/AVX ADDSUB instructions
18811 // during ISel stage.
18812 if (N->getOpcode() == ISD::VECTOR_SHUFFLE &&
18813 ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
18814 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
18815 N0->getOpcode() == ISD::FADD && N1->getOpcode() == ISD::FSUB &&
18816 // Operands to the FADD and FSUB must be the same.
18817 ((N0->getOperand(0) == N1->getOperand(0) &&
18818 N0->getOperand(1) == N1->getOperand(1)) ||
18819 // FADD is commutable. See if by commuting the operands of the FADD
18820 // we would still be able to match the operands of the FSUB dag node.
18821 (N0->getOperand(1) == N1->getOperand(0) &&
18822 N0->getOperand(0) == N1->getOperand(1))) &&
18823 N0->getOperand(0)->getOpcode() != ISD::UNDEF &&
18824 N0->getOperand(1)->getOpcode() != ISD::UNDEF) {
18826 ShuffleVectorSDNode *SV = cast<ShuffleVectorSDNode>(N);
18827 unsigned NumElts = VT.getVectorNumElements();
18828 ArrayRef<int> Mask = SV->getMask();
18829 bool CanFold = true;
18831 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i)
18832 CanFold = Mask[i] == (int)((i & 1) ? i + NumElts : i);
18835 SDValue Op0 = N1->getOperand(0);
18836 SDValue Op1 = DAG.getNode(ISD::FNEG, dl, VT, N1->getOperand(1));
18837 SDValue Sub = DAG.getNode(ISD::FSUB, dl, VT, Op0, Op1);
18838 SDValue Add = DAG.getNode(ISD::FADD, dl, VT, Op0, Op1);
18839 return DAG.getVectorShuffle(VT, dl, Sub, Add, Mask);
18843 // Don't create instructions with illegal types after legalize types has run.
18844 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18845 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
18848 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
18849 if (Subtarget->hasFp256() && VT.is256BitVector() &&
18850 N->getOpcode() == ISD::VECTOR_SHUFFLE)
18851 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
18853 // During Type Legalization, when promoting illegal vector types,
18854 // the backend might introduce new shuffle dag nodes and bitcasts.
18856 // This code performs the following transformation:
18857 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
18858 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
18860 // We do this only if both the bitcast and the BINOP dag nodes have
18861 // one use. Also, perform this transformation only if the new binary
18862 // operation is legal. This is to avoid introducing dag nodes that
18863 // potentially need to be further expanded (or custom lowered) into a
18864 // less optimal sequence of dag nodes.
18865 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
18866 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
18867 N0.getOpcode() == ISD::BITCAST) {
18868 SDValue BC0 = N0.getOperand(0);
18869 EVT SVT = BC0.getValueType();
18870 unsigned Opcode = BC0.getOpcode();
18871 unsigned NumElts = VT.getVectorNumElements();
18873 if (BC0.hasOneUse() && SVT.isVector() &&
18874 SVT.getVectorNumElements() * 2 == NumElts &&
18875 TLI.isOperationLegal(Opcode, VT)) {
18876 bool CanFold = false;
18888 unsigned SVTNumElts = SVT.getVectorNumElements();
18889 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
18890 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
18891 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
18892 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
18893 CanFold = SVOp->getMaskElt(i) < 0;
18896 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
18897 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
18898 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
18899 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
18904 // Only handle 128 wide vector from here on.
18905 if (!VT.is128BitVector())
18908 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
18909 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
18910 // consecutive, non-overlapping, and in the right order.
18911 SmallVector<SDValue, 16> Elts;
18912 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
18913 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
18915 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
18919 if (isTargetShuffle(N->getOpcode())) {
18921 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
18922 if (Shuffle.getNode())
18929 /// PerformTruncateCombine - Converts truncate operation to
18930 /// a sequence of vector shuffle operations.
18931 /// It is possible when we truncate 256-bit vector to 128-bit vector
18932 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
18933 TargetLowering::DAGCombinerInfo &DCI,
18934 const X86Subtarget *Subtarget) {
18938 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
18939 /// specific shuffle of a load can be folded into a single element load.
18940 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
18941 /// shuffles have been customed lowered so we need to handle those here.
18942 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
18943 TargetLowering::DAGCombinerInfo &DCI) {
18944 if (DCI.isBeforeLegalizeOps())
18947 SDValue InVec = N->getOperand(0);
18948 SDValue EltNo = N->getOperand(1);
18950 if (!isa<ConstantSDNode>(EltNo))
18953 EVT VT = InVec.getValueType();
18955 bool HasShuffleIntoBitcast = false;
18956 if (InVec.getOpcode() == ISD::BITCAST) {
18957 // Don't duplicate a load with other uses.
18958 if (!InVec.hasOneUse())
18960 EVT BCVT = InVec.getOperand(0).getValueType();
18961 if (BCVT.getVectorNumElements() != VT.getVectorNumElements())
18963 InVec = InVec.getOperand(0);
18964 HasShuffleIntoBitcast = true;
18967 if (!isTargetShuffle(InVec.getOpcode()))
18970 // Don't duplicate a load with other uses.
18971 if (!InVec.hasOneUse())
18974 SmallVector<int, 16> ShuffleMask;
18976 if (!getTargetShuffleMask(InVec.getNode(), VT.getSimpleVT(), ShuffleMask,
18980 // Select the input vector, guarding against out of range extract vector.
18981 unsigned NumElems = VT.getVectorNumElements();
18982 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
18983 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
18984 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
18985 : InVec.getOperand(1);
18987 // If inputs to shuffle are the same for both ops, then allow 2 uses
18988 unsigned AllowedUses = InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
18990 if (LdNode.getOpcode() == ISD::BITCAST) {
18991 // Don't duplicate a load with other uses.
18992 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
18995 AllowedUses = 1; // only allow 1 load use if we have a bitcast
18996 LdNode = LdNode.getOperand(0);
18999 if (!ISD::isNormalLoad(LdNode.getNode()))
19002 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
19004 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
19007 if (HasShuffleIntoBitcast) {
19008 // If there's a bitcast before the shuffle, check if the load type and
19009 // alignment is valid.
19010 unsigned Align = LN0->getAlignment();
19011 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19012 unsigned NewAlign = TLI.getDataLayout()->
19013 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
19015 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
19019 // All checks match so transform back to vector_shuffle so that DAG combiner
19020 // can finish the job
19023 // Create shuffle node taking into account the case that its a unary shuffle
19024 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(VT) : InVec.getOperand(1);
19025 Shuffle = DAG.getVectorShuffle(InVec.getValueType(), dl,
19026 InVec.getOperand(0), Shuffle,
19028 Shuffle = DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
19029 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
19033 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
19034 /// generation and convert it from being a bunch of shuffles and extracts
19035 /// to a simple store and scalar loads to extract the elements.
19036 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
19037 TargetLowering::DAGCombinerInfo &DCI) {
19038 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
19039 if (NewOp.getNode())
19042 SDValue InputVector = N->getOperand(0);
19044 // Detect whether we are trying to convert from mmx to i32 and the bitcast
19045 // from mmx to v2i32 has a single usage.
19046 if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
19047 InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
19048 InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
19049 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
19050 N->getValueType(0),
19051 InputVector.getNode()->getOperand(0));
19053 // Only operate on vectors of 4 elements, where the alternative shuffling
19054 // gets to be more expensive.
19055 if (InputVector.getValueType() != MVT::v4i32)
19058 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
19059 // single use which is a sign-extend or zero-extend, and all elements are
19061 SmallVector<SDNode *, 4> Uses;
19062 unsigned ExtractedElements = 0;
19063 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
19064 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
19065 if (UI.getUse().getResNo() != InputVector.getResNo())
19068 SDNode *Extract = *UI;
19069 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
19072 if (Extract->getValueType(0) != MVT::i32)
19074 if (!Extract->hasOneUse())
19076 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
19077 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
19079 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
19082 // Record which element was extracted.
19083 ExtractedElements |=
19084 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
19086 Uses.push_back(Extract);
19089 // If not all the elements were used, this may not be worthwhile.
19090 if (ExtractedElements != 15)
19093 // Ok, we've now decided to do the transformation.
19094 SDLoc dl(InputVector);
19096 // Store the value to a temporary stack slot.
19097 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
19098 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
19099 MachinePointerInfo(), false, false, 0);
19101 // Replace each use (extract) with a load of the appropriate element.
19102 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
19103 UE = Uses.end(); UI != UE; ++UI) {
19104 SDNode *Extract = *UI;
19106 // cOMpute the element's address.
19107 SDValue Idx = Extract->getOperand(1);
19109 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
19110 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
19111 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19112 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
19114 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
19115 StackPtr, OffsetVal);
19117 // Load the scalar.
19118 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
19119 ScalarAddr, MachinePointerInfo(),
19120 false, false, false, 0);
19122 // Replace the exact with the load.
19123 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
19126 // The replacement was made in place; don't return anything.
19130 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
19131 static std::pair<unsigned, bool>
19132 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
19133 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
19134 if (!VT.isVector())
19135 return std::make_pair(0, false);
19137 bool NeedSplit = false;
19138 switch (VT.getSimpleVT().SimpleTy) {
19139 default: return std::make_pair(0, false);
19143 if (!Subtarget->hasAVX2())
19145 if (!Subtarget->hasAVX())
19146 return std::make_pair(0, false);
19151 if (!Subtarget->hasSSE2())
19152 return std::make_pair(0, false);
19155 // SSE2 has only a small subset of the operations.
19156 bool hasUnsigned = Subtarget->hasSSE41() ||
19157 (Subtarget->hasSSE2() && VT == MVT::v16i8);
19158 bool hasSigned = Subtarget->hasSSE41() ||
19159 (Subtarget->hasSSE2() && VT == MVT::v8i16);
19161 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
19164 // Check for x CC y ? x : y.
19165 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
19166 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
19171 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
19174 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
19177 Opc = hasSigned ? X86ISD::SMIN : 0; break;
19180 Opc = hasSigned ? X86ISD::SMAX : 0; break;
19182 // Check for x CC y ? y : x -- a min/max with reversed arms.
19183 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
19184 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
19189 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
19192 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
19195 Opc = hasSigned ? X86ISD::SMAX : 0; break;
19198 Opc = hasSigned ? X86ISD::SMIN : 0; break;
19202 return std::make_pair(Opc, NeedSplit);
19206 TransformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
19207 const X86Subtarget *Subtarget) {
19209 SDValue Cond = N->getOperand(0);
19210 SDValue LHS = N->getOperand(1);
19211 SDValue RHS = N->getOperand(2);
19213 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
19214 SDValue CondSrc = Cond->getOperand(0);
19215 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
19216 Cond = CondSrc->getOperand(0);
19219 MVT VT = N->getSimpleValueType(0);
19220 MVT EltVT = VT.getVectorElementType();
19221 unsigned NumElems = VT.getVectorNumElements();
19222 // There is no blend with immediate in AVX-512.
19223 if (VT.is512BitVector())
19226 if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
19228 if (!Subtarget->hasInt256() && VT == MVT::v16i16)
19231 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
19234 unsigned MaskValue = 0;
19235 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
19238 SmallVector<int, 8> ShuffleMask(NumElems, -1);
19239 for (unsigned i = 0; i < NumElems; ++i) {
19240 // Be sure we emit undef where we can.
19241 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
19242 ShuffleMask[i] = -1;
19244 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
19247 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
19250 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
19252 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
19253 TargetLowering::DAGCombinerInfo &DCI,
19254 const X86Subtarget *Subtarget) {
19256 SDValue Cond = N->getOperand(0);
19257 // Get the LHS/RHS of the select.
19258 SDValue LHS = N->getOperand(1);
19259 SDValue RHS = N->getOperand(2);
19260 EVT VT = LHS.getValueType();
19261 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19263 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
19264 // instructions match the semantics of the common C idiom x<y?x:y but not
19265 // x<=y?x:y, because of how they handle negative zero (which can be
19266 // ignored in unsafe-math mode).
19267 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
19268 VT != MVT::f80 && TLI.isTypeLegal(VT) &&
19269 (Subtarget->hasSSE2() ||
19270 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
19271 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
19273 unsigned Opcode = 0;
19274 // Check for x CC y ? x : y.
19275 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
19276 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
19280 // Converting this to a min would handle NaNs incorrectly, and swapping
19281 // the operands would cause it to handle comparisons between positive
19282 // and negative zero incorrectly.
19283 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
19284 if (!DAG.getTarget().Options.UnsafeFPMath &&
19285 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
19287 std::swap(LHS, RHS);
19289 Opcode = X86ISD::FMIN;
19292 // Converting this to a min would handle comparisons between positive
19293 // and negative zero incorrectly.
19294 if (!DAG.getTarget().Options.UnsafeFPMath &&
19295 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
19297 Opcode = X86ISD::FMIN;
19300 // Converting this to a min would handle both negative zeros and NaNs
19301 // incorrectly, but we can swap the operands to fix both.
19302 std::swap(LHS, RHS);
19306 Opcode = X86ISD::FMIN;
19310 // Converting this to a max would handle comparisons between positive
19311 // and negative zero incorrectly.
19312 if (!DAG.getTarget().Options.UnsafeFPMath &&
19313 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
19315 Opcode = X86ISD::FMAX;
19318 // Converting this to a max would handle NaNs incorrectly, and swapping
19319 // the operands would cause it to handle comparisons between positive
19320 // and negative zero incorrectly.
19321 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
19322 if (!DAG.getTarget().Options.UnsafeFPMath &&
19323 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
19325 std::swap(LHS, RHS);
19327 Opcode = X86ISD::FMAX;
19330 // Converting this to a max would handle both negative zeros and NaNs
19331 // incorrectly, but we can swap the operands to fix both.
19332 std::swap(LHS, RHS);
19336 Opcode = X86ISD::FMAX;
19339 // Check for x CC y ? y : x -- a min/max with reversed arms.
19340 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
19341 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
19345 // Converting this to a min would handle comparisons between positive
19346 // and negative zero incorrectly, and swapping the operands would
19347 // cause it to handle NaNs incorrectly.
19348 if (!DAG.getTarget().Options.UnsafeFPMath &&
19349 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
19350 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
19352 std::swap(LHS, RHS);
19354 Opcode = X86ISD::FMIN;
19357 // Converting this to a min would handle NaNs incorrectly.
19358 if (!DAG.getTarget().Options.UnsafeFPMath &&
19359 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
19361 Opcode = X86ISD::FMIN;
19364 // Converting this to a min would handle both negative zeros and NaNs
19365 // incorrectly, but we can swap the operands to fix both.
19366 std::swap(LHS, RHS);
19370 Opcode = X86ISD::FMIN;
19374 // Converting this to a max would handle NaNs incorrectly.
19375 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
19377 Opcode = X86ISD::FMAX;
19380 // Converting this to a max would handle comparisons between positive
19381 // and negative zero incorrectly, and swapping the operands would
19382 // cause it to handle NaNs incorrectly.
19383 if (!DAG.getTarget().Options.UnsafeFPMath &&
19384 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
19385 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
19387 std::swap(LHS, RHS);
19389 Opcode = X86ISD::FMAX;
19392 // Converting this to a max would handle both negative zeros and NaNs
19393 // incorrectly, but we can swap the operands to fix both.
19394 std::swap(LHS, RHS);
19398 Opcode = X86ISD::FMAX;
19404 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
19407 EVT CondVT = Cond.getValueType();
19408 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
19409 CondVT.getVectorElementType() == MVT::i1) {
19410 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
19411 // lowering on AVX-512. In this case we convert it to
19412 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
19413 // The same situation for all 128 and 256-bit vectors of i8 and i16
19414 EVT OpVT = LHS.getValueType();
19415 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
19416 (OpVT.getVectorElementType() == MVT::i8 ||
19417 OpVT.getVectorElementType() == MVT::i16)) {
19418 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
19419 DCI.AddToWorklist(Cond.getNode());
19420 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
19423 // If this is a select between two integer constants, try to do some
19425 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
19426 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
19427 // Don't do this for crazy integer types.
19428 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
19429 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
19430 // so that TrueC (the true value) is larger than FalseC.
19431 bool NeedsCondInvert = false;
19433 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
19434 // Efficiently invertible.
19435 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
19436 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
19437 isa<ConstantSDNode>(Cond.getOperand(1))))) {
19438 NeedsCondInvert = true;
19439 std::swap(TrueC, FalseC);
19442 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
19443 if (FalseC->getAPIntValue() == 0 &&
19444 TrueC->getAPIntValue().isPowerOf2()) {
19445 if (NeedsCondInvert) // Invert the condition if needed.
19446 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
19447 DAG.getConstant(1, Cond.getValueType()));
19449 // Zero extend the condition if needed.
19450 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
19452 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
19453 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
19454 DAG.getConstant(ShAmt, MVT::i8));
19457 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
19458 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
19459 if (NeedsCondInvert) // Invert the condition if needed.
19460 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
19461 DAG.getConstant(1, Cond.getValueType()));
19463 // Zero extend the condition if needed.
19464 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
19465 FalseC->getValueType(0), Cond);
19466 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
19467 SDValue(FalseC, 0));
19470 // Optimize cases that will turn into an LEA instruction. This requires
19471 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
19472 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
19473 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
19474 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
19476 bool isFastMultiplier = false;
19478 switch ((unsigned char)Diff) {
19480 case 1: // result = add base, cond
19481 case 2: // result = lea base( , cond*2)
19482 case 3: // result = lea base(cond, cond*2)
19483 case 4: // result = lea base( , cond*4)
19484 case 5: // result = lea base(cond, cond*4)
19485 case 8: // result = lea base( , cond*8)
19486 case 9: // result = lea base(cond, cond*8)
19487 isFastMultiplier = true;
19492 if (isFastMultiplier) {
19493 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
19494 if (NeedsCondInvert) // Invert the condition if needed.
19495 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
19496 DAG.getConstant(1, Cond.getValueType()));
19498 // Zero extend the condition if needed.
19499 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
19501 // Scale the condition by the difference.
19503 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
19504 DAG.getConstant(Diff, Cond.getValueType()));
19506 // Add the base if non-zero.
19507 if (FalseC->getAPIntValue() != 0)
19508 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
19509 SDValue(FalseC, 0));
19516 // Canonicalize max and min:
19517 // (x > y) ? x : y -> (x >= y) ? x : y
19518 // (x < y) ? x : y -> (x <= y) ? x : y
19519 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
19520 // the need for an extra compare
19521 // against zero. e.g.
19522 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
19524 // testl %edi, %edi
19526 // cmovgl %edi, %eax
19530 // cmovsl %eax, %edi
19531 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
19532 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
19533 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
19534 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
19539 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
19540 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
19541 Cond.getOperand(0), Cond.getOperand(1), NewCC);
19542 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
19547 // Early exit check
19548 if (!TLI.isTypeLegal(VT))
19551 // Match VSELECTs into subs with unsigned saturation.
19552 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
19553 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
19554 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
19555 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
19556 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
19558 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
19559 // left side invert the predicate to simplify logic below.
19561 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
19563 CC = ISD::getSetCCInverse(CC, true);
19564 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
19568 if (Other.getNode() && Other->getNumOperands() == 2 &&
19569 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
19570 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
19571 SDValue CondRHS = Cond->getOperand(1);
19573 // Look for a general sub with unsigned saturation first.
19574 // x >= y ? x-y : 0 --> subus x, y
19575 // x > y ? x-y : 0 --> subus x, y
19576 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
19577 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
19578 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
19580 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
19581 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
19582 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
19583 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
19584 // If the RHS is a constant we have to reverse the const
19585 // canonicalization.
19586 // x > C-1 ? x+-C : 0 --> subus x, C
19587 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
19588 CondRHSConst->getAPIntValue() ==
19589 (-OpRHSConst->getAPIntValue() - 1))
19590 return DAG.getNode(
19591 X86ISD::SUBUS, DL, VT, OpLHS,
19592 DAG.getConstant(-OpRHSConst->getAPIntValue(), VT));
19594 // Another special case: If C was a sign bit, the sub has been
19595 // canonicalized into a xor.
19596 // FIXME: Would it be better to use computeKnownBits to determine
19597 // whether it's safe to decanonicalize the xor?
19598 // x s< 0 ? x^C : 0 --> subus x, C
19599 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
19600 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
19601 OpRHSConst->getAPIntValue().isSignBit())
19602 // Note that we have to rebuild the RHS constant here to ensure we
19603 // don't rely on particular values of undef lanes.
19604 return DAG.getNode(
19605 X86ISD::SUBUS, DL, VT, OpLHS,
19606 DAG.getConstant(OpRHSConst->getAPIntValue(), VT));
19611 // Try to match a min/max vector operation.
19612 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
19613 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
19614 unsigned Opc = ret.first;
19615 bool NeedSplit = ret.second;
19617 if (Opc && NeedSplit) {
19618 unsigned NumElems = VT.getVectorNumElements();
19619 // Extract the LHS vectors
19620 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
19621 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
19623 // Extract the RHS vectors
19624 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
19625 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
19627 // Create min/max for each subvector
19628 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
19629 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
19631 // Merge the result
19632 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
19634 return DAG.getNode(Opc, DL, VT, LHS, RHS);
19637 // Simplify vector selection if the selector will be produced by CMPP*/PCMP*.
19638 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
19639 // Check if SETCC has already been promoted
19640 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT &&
19641 // Check that condition value type matches vselect operand type
19644 assert(Cond.getValueType().isVector() &&
19645 "vector select expects a vector selector!");
19647 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
19648 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
19650 if (!TValIsAllOnes && !FValIsAllZeros) {
19651 // Try invert the condition if true value is not all 1s and false value
19653 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
19654 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
19656 if (TValIsAllZeros || FValIsAllOnes) {
19657 SDValue CC = Cond.getOperand(2);
19658 ISD::CondCode NewCC =
19659 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
19660 Cond.getOperand(0).getValueType().isInteger());
19661 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
19662 std::swap(LHS, RHS);
19663 TValIsAllOnes = FValIsAllOnes;
19664 FValIsAllZeros = TValIsAllZeros;
19668 if (TValIsAllOnes || FValIsAllZeros) {
19671 if (TValIsAllOnes && FValIsAllZeros)
19673 else if (TValIsAllOnes)
19674 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
19675 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
19676 else if (FValIsAllZeros)
19677 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
19678 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
19680 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
19684 // Try to fold this VSELECT into a MOVSS/MOVSD
19685 if (N->getOpcode() == ISD::VSELECT &&
19686 Cond.getOpcode() == ISD::BUILD_VECTOR && !DCI.isBeforeLegalize()) {
19687 if (VT == MVT::v4i32 || VT == MVT::v4f32 ||
19688 (Subtarget->hasSSE2() && (VT == MVT::v2i64 || VT == MVT::v2f64))) {
19689 bool CanFold = false;
19690 unsigned NumElems = Cond.getNumOperands();
19694 if (isZero(Cond.getOperand(0))) {
19697 // fold (vselect <0,-1,-1,-1>, A, B) -> (movss A, B)
19698 // fold (vselect <0,-1> -> (movsd A, B)
19699 for (unsigned i = 1, e = NumElems; i != e && CanFold; ++i)
19700 CanFold = isAllOnes(Cond.getOperand(i));
19701 } else if (isAllOnes(Cond.getOperand(0))) {
19705 // fold (vselect <-1,0,0,0>, A, B) -> (movss B, A)
19706 // fold (vselect <-1,0> -> (movsd B, A)
19707 for (unsigned i = 1, e = NumElems; i != e && CanFold; ++i)
19708 CanFold = isZero(Cond.getOperand(i));
19712 if (VT == MVT::v4i32 || VT == MVT::v4f32)
19713 return getTargetShuffleNode(X86ISD::MOVSS, DL, VT, A, B, DAG);
19714 return getTargetShuffleNode(X86ISD::MOVSD, DL, VT, A, B, DAG);
19717 if (Subtarget->hasSSE2() && (VT == MVT::v4i32 || VT == MVT::v4f32)) {
19718 // fold (v4i32: vselect <0,0,-1,-1>, A, B) ->
19719 // (v4i32 (bitcast (movsd (v2i64 (bitcast A)),
19720 // (v2i64 (bitcast B)))))
19722 // fold (v4f32: vselect <0,0,-1,-1>, A, B) ->
19723 // (v4f32 (bitcast (movsd (v2f64 (bitcast A)),
19724 // (v2f64 (bitcast B)))))
19726 // fold (v4i32: vselect <-1,-1,0,0>, A, B) ->
19727 // (v4i32 (bitcast (movsd (v2i64 (bitcast B)),
19728 // (v2i64 (bitcast A)))))
19730 // fold (v4f32: vselect <-1,-1,0,0>, A, B) ->
19731 // (v4f32 (bitcast (movsd (v2f64 (bitcast B)),
19732 // (v2f64 (bitcast A)))))
19734 CanFold = (isZero(Cond.getOperand(0)) &&
19735 isZero(Cond.getOperand(1)) &&
19736 isAllOnes(Cond.getOperand(2)) &&
19737 isAllOnes(Cond.getOperand(3)));
19739 if (!CanFold && isAllOnes(Cond.getOperand(0)) &&
19740 isAllOnes(Cond.getOperand(1)) &&
19741 isZero(Cond.getOperand(2)) &&
19742 isZero(Cond.getOperand(3))) {
19744 std::swap(LHS, RHS);
19748 EVT NVT = (VT == MVT::v4i32) ? MVT::v2i64 : MVT::v2f64;
19749 SDValue NewA = DAG.getNode(ISD::BITCAST, DL, NVT, LHS);
19750 SDValue NewB = DAG.getNode(ISD::BITCAST, DL, NVT, RHS);
19751 SDValue Select = getTargetShuffleNode(X86ISD::MOVSD, DL, NVT, NewA,
19753 return DAG.getNode(ISD::BITCAST, DL, VT, Select);
19759 // If we know that this node is legal then we know that it is going to be
19760 // matched by one of the SSE/AVX BLEND instructions. These instructions only
19761 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
19762 // to simplify previous instructions.
19763 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
19764 !DCI.isBeforeLegalize() &&
19765 // We explicitly check against v8i16 and v16i16 because, although
19766 // they're marked as Custom, they might only be legal when Cond is a
19767 // build_vector of constants. This will be taken care in a later
19769 (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) && VT != MVT::v16i16 &&
19770 VT != MVT::v8i16)) {
19771 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
19773 // Don't optimize vector selects that map to mask-registers.
19777 // Check all uses of that condition operand to check whether it will be
19778 // consumed by non-BLEND instructions, which may depend on all bits are set
19780 for (SDNode::use_iterator I = Cond->use_begin(),
19781 E = Cond->use_end(); I != E; ++I)
19782 if (I->getOpcode() != ISD::VSELECT)
19783 // TODO: Add other opcodes eventually lowered into BLEND.
19786 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
19787 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
19789 APInt KnownZero, KnownOne;
19790 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
19791 DCI.isBeforeLegalizeOps());
19792 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
19793 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne, TLO))
19794 DCI.CommitTargetLoweringOpt(TLO);
19797 // We should generate an X86ISD::BLENDI from a vselect if its argument
19798 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
19799 // constants. This specific pattern gets generated when we split a
19800 // selector for a 512 bit vector in a machine without AVX512 (but with
19801 // 256-bit vectors), during legalization:
19803 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
19805 // Iff we find this pattern and the build_vectors are built from
19806 // constants, we translate the vselect into a shuffle_vector that we
19807 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
19808 if (N->getOpcode() == ISD::VSELECT && !DCI.isBeforeLegalize()) {
19809 SDValue Shuffle = TransformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
19810 if (Shuffle.getNode())
19817 // Check whether a boolean test is testing a boolean value generated by
19818 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
19821 // Simplify the following patterns:
19822 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
19823 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
19824 // to (Op EFLAGS Cond)
19826 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
19827 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
19828 // to (Op EFLAGS !Cond)
19830 // where Op could be BRCOND or CMOV.
19832 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
19833 // Quit if not CMP and SUB with its value result used.
19834 if (Cmp.getOpcode() != X86ISD::CMP &&
19835 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
19838 // Quit if not used as a boolean value.
19839 if (CC != X86::COND_E && CC != X86::COND_NE)
19842 // Check CMP operands. One of them should be 0 or 1 and the other should be
19843 // an SetCC or extended from it.
19844 SDValue Op1 = Cmp.getOperand(0);
19845 SDValue Op2 = Cmp.getOperand(1);
19848 const ConstantSDNode* C = nullptr;
19849 bool needOppositeCond = (CC == X86::COND_E);
19850 bool checkAgainstTrue = false; // Is it a comparison against 1?
19852 if ((C = dyn_cast<ConstantSDNode>(Op1)))
19854 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
19856 else // Quit if all operands are not constants.
19859 if (C->getZExtValue() == 1) {
19860 needOppositeCond = !needOppositeCond;
19861 checkAgainstTrue = true;
19862 } else if (C->getZExtValue() != 0)
19863 // Quit if the constant is neither 0 or 1.
19866 bool truncatedToBoolWithAnd = false;
19867 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
19868 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
19869 SetCC.getOpcode() == ISD::TRUNCATE ||
19870 SetCC.getOpcode() == ISD::AND) {
19871 if (SetCC.getOpcode() == ISD::AND) {
19873 ConstantSDNode *CS;
19874 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
19875 CS->getZExtValue() == 1)
19877 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
19878 CS->getZExtValue() == 1)
19882 SetCC = SetCC.getOperand(OpIdx);
19883 truncatedToBoolWithAnd = true;
19885 SetCC = SetCC.getOperand(0);
19888 switch (SetCC.getOpcode()) {
19889 case X86ISD::SETCC_CARRY:
19890 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
19891 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
19892 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
19893 // truncated to i1 using 'and'.
19894 if (checkAgainstTrue && !truncatedToBoolWithAnd)
19896 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
19897 "Invalid use of SETCC_CARRY!");
19899 case X86ISD::SETCC:
19900 // Set the condition code or opposite one if necessary.
19901 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
19902 if (needOppositeCond)
19903 CC = X86::GetOppositeBranchCondition(CC);
19904 return SetCC.getOperand(1);
19905 case X86ISD::CMOV: {
19906 // Check whether false/true value has canonical one, i.e. 0 or 1.
19907 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
19908 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
19909 // Quit if true value is not a constant.
19912 // Quit if false value is not a constant.
19914 SDValue Op = SetCC.getOperand(0);
19915 // Skip 'zext' or 'trunc' node.
19916 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
19917 Op.getOpcode() == ISD::TRUNCATE)
19918 Op = Op.getOperand(0);
19919 // A special case for rdrand/rdseed, where 0 is set if false cond is
19921 if ((Op.getOpcode() != X86ISD::RDRAND &&
19922 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
19925 // Quit if false value is not the constant 0 or 1.
19926 bool FValIsFalse = true;
19927 if (FVal && FVal->getZExtValue() != 0) {
19928 if (FVal->getZExtValue() != 1)
19930 // If FVal is 1, opposite cond is needed.
19931 needOppositeCond = !needOppositeCond;
19932 FValIsFalse = false;
19934 // Quit if TVal is not the constant opposite of FVal.
19935 if (FValIsFalse && TVal->getZExtValue() != 1)
19937 if (!FValIsFalse && TVal->getZExtValue() != 0)
19939 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
19940 if (needOppositeCond)
19941 CC = X86::GetOppositeBranchCondition(CC);
19942 return SetCC.getOperand(3);
19949 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
19950 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
19951 TargetLowering::DAGCombinerInfo &DCI,
19952 const X86Subtarget *Subtarget) {
19955 // If the flag operand isn't dead, don't touch this CMOV.
19956 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
19959 SDValue FalseOp = N->getOperand(0);
19960 SDValue TrueOp = N->getOperand(1);
19961 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
19962 SDValue Cond = N->getOperand(3);
19964 if (CC == X86::COND_E || CC == X86::COND_NE) {
19965 switch (Cond.getOpcode()) {
19969 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
19970 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
19971 return (CC == X86::COND_E) ? FalseOp : TrueOp;
19977 Flags = checkBoolTestSetCCCombine(Cond, CC);
19978 if (Flags.getNode() &&
19979 // Extra check as FCMOV only supports a subset of X86 cond.
19980 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
19981 SDValue Ops[] = { FalseOp, TrueOp,
19982 DAG.getConstant(CC, MVT::i8), Flags };
19983 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
19986 // If this is a select between two integer constants, try to do some
19987 // optimizations. Note that the operands are ordered the opposite of SELECT
19989 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
19990 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
19991 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
19992 // larger than FalseC (the false value).
19993 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
19994 CC = X86::GetOppositeBranchCondition(CC);
19995 std::swap(TrueC, FalseC);
19996 std::swap(TrueOp, FalseOp);
19999 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
20000 // This is efficient for any integer data type (including i8/i16) and
20002 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
20003 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
20004 DAG.getConstant(CC, MVT::i8), Cond);
20006 // Zero extend the condition if needed.
20007 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
20009 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
20010 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
20011 DAG.getConstant(ShAmt, MVT::i8));
20012 if (N->getNumValues() == 2) // Dead flag value?
20013 return DCI.CombineTo(N, Cond, SDValue());
20017 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
20018 // for any integer data type, including i8/i16.
20019 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
20020 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
20021 DAG.getConstant(CC, MVT::i8), Cond);
20023 // Zero extend the condition if needed.
20024 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
20025 FalseC->getValueType(0), Cond);
20026 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
20027 SDValue(FalseC, 0));
20029 if (N->getNumValues() == 2) // Dead flag value?
20030 return DCI.CombineTo(N, Cond, SDValue());
20034 // Optimize cases that will turn into an LEA instruction. This requires
20035 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
20036 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
20037 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
20038 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
20040 bool isFastMultiplier = false;
20042 switch ((unsigned char)Diff) {
20044 case 1: // result = add base, cond
20045 case 2: // result = lea base( , cond*2)
20046 case 3: // result = lea base(cond, cond*2)
20047 case 4: // result = lea base( , cond*4)
20048 case 5: // result = lea base(cond, cond*4)
20049 case 8: // result = lea base( , cond*8)
20050 case 9: // result = lea base(cond, cond*8)
20051 isFastMultiplier = true;
20056 if (isFastMultiplier) {
20057 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
20058 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
20059 DAG.getConstant(CC, MVT::i8), Cond);
20060 // Zero extend the condition if needed.
20061 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
20063 // Scale the condition by the difference.
20065 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
20066 DAG.getConstant(Diff, Cond.getValueType()));
20068 // Add the base if non-zero.
20069 if (FalseC->getAPIntValue() != 0)
20070 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
20071 SDValue(FalseC, 0));
20072 if (N->getNumValues() == 2) // Dead flag value?
20073 return DCI.CombineTo(N, Cond, SDValue());
20080 // Handle these cases:
20081 // (select (x != c), e, c) -> select (x != c), e, x),
20082 // (select (x == c), c, e) -> select (x == c), x, e)
20083 // where the c is an integer constant, and the "select" is the combination
20084 // of CMOV and CMP.
20086 // The rationale for this change is that the conditional-move from a constant
20087 // needs two instructions, however, conditional-move from a register needs
20088 // only one instruction.
20090 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
20091 // some instruction-combining opportunities. This opt needs to be
20092 // postponed as late as possible.
20094 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
20095 // the DCI.xxxx conditions are provided to postpone the optimization as
20096 // late as possible.
20098 ConstantSDNode *CmpAgainst = nullptr;
20099 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
20100 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
20101 !isa<ConstantSDNode>(Cond.getOperand(0))) {
20103 if (CC == X86::COND_NE &&
20104 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
20105 CC = X86::GetOppositeBranchCondition(CC);
20106 std::swap(TrueOp, FalseOp);
20109 if (CC == X86::COND_E &&
20110 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
20111 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
20112 DAG.getConstant(CC, MVT::i8), Cond };
20113 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
20121 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
20122 const X86Subtarget *Subtarget) {
20123 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
20125 default: return SDValue();
20126 // SSE/AVX/AVX2 blend intrinsics.
20127 case Intrinsic::x86_avx2_pblendvb:
20128 case Intrinsic::x86_avx2_pblendw:
20129 case Intrinsic::x86_avx2_pblendd_128:
20130 case Intrinsic::x86_avx2_pblendd_256:
20131 // Don't try to simplify this intrinsic if we don't have AVX2.
20132 if (!Subtarget->hasAVX2())
20135 case Intrinsic::x86_avx_blend_pd_256:
20136 case Intrinsic::x86_avx_blend_ps_256:
20137 case Intrinsic::x86_avx_blendv_pd_256:
20138 case Intrinsic::x86_avx_blendv_ps_256:
20139 // Don't try to simplify this intrinsic if we don't have AVX.
20140 if (!Subtarget->hasAVX())
20143 case Intrinsic::x86_sse41_pblendw:
20144 case Intrinsic::x86_sse41_blendpd:
20145 case Intrinsic::x86_sse41_blendps:
20146 case Intrinsic::x86_sse41_blendvps:
20147 case Intrinsic::x86_sse41_blendvpd:
20148 case Intrinsic::x86_sse41_pblendvb: {
20149 SDValue Op0 = N->getOperand(1);
20150 SDValue Op1 = N->getOperand(2);
20151 SDValue Mask = N->getOperand(3);
20153 // Don't try to simplify this intrinsic if we don't have SSE4.1.
20154 if (!Subtarget->hasSSE41())
20157 // fold (blend A, A, Mask) -> A
20160 // fold (blend A, B, allZeros) -> A
20161 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
20163 // fold (blend A, B, allOnes) -> B
20164 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
20167 // Simplify the case where the mask is a constant i32 value.
20168 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
20169 if (C->isNullValue())
20171 if (C->isAllOnesValue())
20178 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
20179 case Intrinsic::x86_sse2_psrai_w:
20180 case Intrinsic::x86_sse2_psrai_d:
20181 case Intrinsic::x86_avx2_psrai_w:
20182 case Intrinsic::x86_avx2_psrai_d:
20183 case Intrinsic::x86_sse2_psra_w:
20184 case Intrinsic::x86_sse2_psra_d:
20185 case Intrinsic::x86_avx2_psra_w:
20186 case Intrinsic::x86_avx2_psra_d: {
20187 SDValue Op0 = N->getOperand(1);
20188 SDValue Op1 = N->getOperand(2);
20189 EVT VT = Op0.getValueType();
20190 assert(VT.isVector() && "Expected a vector type!");
20192 if (isa<BuildVectorSDNode>(Op1))
20193 Op1 = Op1.getOperand(0);
20195 if (!isa<ConstantSDNode>(Op1))
20198 EVT SVT = VT.getVectorElementType();
20199 unsigned SVTBits = SVT.getSizeInBits();
20201 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
20202 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
20203 uint64_t ShAmt = C.getZExtValue();
20205 // Don't try to convert this shift into a ISD::SRA if the shift
20206 // count is bigger than or equal to the element size.
20207 if (ShAmt >= SVTBits)
20210 // Trivial case: if the shift count is zero, then fold this
20211 // into the first operand.
20215 // Replace this packed shift intrinsic with a target independent
20217 SDValue Splat = DAG.getConstant(C, VT);
20218 return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat);
20223 /// PerformMulCombine - Optimize a single multiply with constant into two
20224 /// in order to implement it with two cheaper instructions, e.g.
20225 /// LEA + SHL, LEA + LEA.
20226 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
20227 TargetLowering::DAGCombinerInfo &DCI) {
20228 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
20231 EVT VT = N->getValueType(0);
20232 if (VT != MVT::i64)
20235 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
20238 uint64_t MulAmt = C->getZExtValue();
20239 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
20242 uint64_t MulAmt1 = 0;
20243 uint64_t MulAmt2 = 0;
20244 if ((MulAmt % 9) == 0) {
20246 MulAmt2 = MulAmt / 9;
20247 } else if ((MulAmt % 5) == 0) {
20249 MulAmt2 = MulAmt / 5;
20250 } else if ((MulAmt % 3) == 0) {
20252 MulAmt2 = MulAmt / 3;
20255 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
20258 if (isPowerOf2_64(MulAmt2) &&
20259 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
20260 // If second multiplifer is pow2, issue it first. We want the multiply by
20261 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
20263 std::swap(MulAmt1, MulAmt2);
20266 if (isPowerOf2_64(MulAmt1))
20267 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
20268 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
20270 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
20271 DAG.getConstant(MulAmt1, VT));
20273 if (isPowerOf2_64(MulAmt2))
20274 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
20275 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
20277 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
20278 DAG.getConstant(MulAmt2, VT));
20280 // Do not add new nodes to DAG combiner worklist.
20281 DCI.CombineTo(N, NewMul, false);
20286 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
20287 SDValue N0 = N->getOperand(0);
20288 SDValue N1 = N->getOperand(1);
20289 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
20290 EVT VT = N0.getValueType();
20292 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
20293 // since the result of setcc_c is all zero's or all ones.
20294 if (VT.isInteger() && !VT.isVector() &&
20295 N1C && N0.getOpcode() == ISD::AND &&
20296 N0.getOperand(1).getOpcode() == ISD::Constant) {
20297 SDValue N00 = N0.getOperand(0);
20298 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
20299 ((N00.getOpcode() == ISD::ANY_EXTEND ||
20300 N00.getOpcode() == ISD::ZERO_EXTEND) &&
20301 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
20302 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
20303 APInt ShAmt = N1C->getAPIntValue();
20304 Mask = Mask.shl(ShAmt);
20306 return DAG.getNode(ISD::AND, SDLoc(N), VT,
20307 N00, DAG.getConstant(Mask, VT));
20311 // Hardware support for vector shifts is sparse which makes us scalarize the
20312 // vector operations in many cases. Also, on sandybridge ADD is faster than
20314 // (shl V, 1) -> add V,V
20315 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
20316 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
20317 assert(N0.getValueType().isVector() && "Invalid vector shift type");
20318 // We shift all of the values by one. In many cases we do not have
20319 // hardware support for this operation. This is better expressed as an ADD
20321 if (N1SplatC->getZExtValue() == 1)
20322 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
20328 /// \brief Returns a vector of 0s if the node in input is a vector logical
20329 /// shift by a constant amount which is known to be bigger than or equal
20330 /// to the vector element size in bits.
20331 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
20332 const X86Subtarget *Subtarget) {
20333 EVT VT = N->getValueType(0);
20335 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
20336 (!Subtarget->hasInt256() ||
20337 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
20340 SDValue Amt = N->getOperand(1);
20342 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
20343 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
20344 APInt ShiftAmt = AmtSplat->getAPIntValue();
20345 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
20347 // SSE2/AVX2 logical shifts always return a vector of 0s
20348 // if the shift amount is bigger than or equal to
20349 // the element size. The constant shift amount will be
20350 // encoded as a 8-bit immediate.
20351 if (ShiftAmt.trunc(8).uge(MaxAmount))
20352 return getZeroVector(VT, Subtarget, DAG, DL);
20358 /// PerformShiftCombine - Combine shifts.
20359 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
20360 TargetLowering::DAGCombinerInfo &DCI,
20361 const X86Subtarget *Subtarget) {
20362 if (N->getOpcode() == ISD::SHL) {
20363 SDValue V = PerformSHLCombine(N, DAG);
20364 if (V.getNode()) return V;
20367 if (N->getOpcode() != ISD::SRA) {
20368 // Try to fold this logical shift into a zero vector.
20369 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
20370 if (V.getNode()) return V;
20376 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
20377 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
20378 // and friends. Likewise for OR -> CMPNEQSS.
20379 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
20380 TargetLowering::DAGCombinerInfo &DCI,
20381 const X86Subtarget *Subtarget) {
20384 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
20385 // we're requiring SSE2 for both.
20386 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
20387 SDValue N0 = N->getOperand(0);
20388 SDValue N1 = N->getOperand(1);
20389 SDValue CMP0 = N0->getOperand(1);
20390 SDValue CMP1 = N1->getOperand(1);
20393 // The SETCCs should both refer to the same CMP.
20394 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
20397 SDValue CMP00 = CMP0->getOperand(0);
20398 SDValue CMP01 = CMP0->getOperand(1);
20399 EVT VT = CMP00.getValueType();
20401 if (VT == MVT::f32 || VT == MVT::f64) {
20402 bool ExpectingFlags = false;
20403 // Check for any users that want flags:
20404 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
20405 !ExpectingFlags && UI != UE; ++UI)
20406 switch (UI->getOpcode()) {
20411 ExpectingFlags = true;
20413 case ISD::CopyToReg:
20414 case ISD::SIGN_EXTEND:
20415 case ISD::ZERO_EXTEND:
20416 case ISD::ANY_EXTEND:
20420 if (!ExpectingFlags) {
20421 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
20422 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
20424 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
20425 X86::CondCode tmp = cc0;
20430 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
20431 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
20432 // FIXME: need symbolic constants for these magic numbers.
20433 // See X86ATTInstPrinter.cpp:printSSECC().
20434 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
20435 if (Subtarget->hasAVX512()) {
20436 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
20437 CMP01, DAG.getConstant(x86cc, MVT::i8));
20438 if (N->getValueType(0) != MVT::i1)
20439 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
20443 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
20444 CMP00.getValueType(), CMP00, CMP01,
20445 DAG.getConstant(x86cc, MVT::i8));
20447 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
20448 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
20450 if (is64BitFP && !Subtarget->is64Bit()) {
20451 // On a 32-bit target, we cannot bitcast the 64-bit float to a
20452 // 64-bit integer, since that's not a legal type. Since
20453 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
20454 // bits, but can do this little dance to extract the lowest 32 bits
20455 // and work with those going forward.
20456 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
20458 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
20460 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
20461 Vector32, DAG.getIntPtrConstant(0));
20465 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
20466 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
20467 DAG.getConstant(1, IntVT));
20468 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
20469 return OneBitOfTruth;
20477 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
20478 /// so it can be folded inside ANDNP.
20479 static bool CanFoldXORWithAllOnes(const SDNode *N) {
20480 EVT VT = N->getValueType(0);
20482 // Match direct AllOnes for 128 and 256-bit vectors
20483 if (ISD::isBuildVectorAllOnes(N))
20486 // Look through a bit convert.
20487 if (N->getOpcode() == ISD::BITCAST)
20488 N = N->getOperand(0).getNode();
20490 // Sometimes the operand may come from a insert_subvector building a 256-bit
20492 if (VT.is256BitVector() &&
20493 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
20494 SDValue V1 = N->getOperand(0);
20495 SDValue V2 = N->getOperand(1);
20497 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
20498 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
20499 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
20500 ISD::isBuildVectorAllOnes(V2.getNode()))
20507 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
20508 // register. In most cases we actually compare or select YMM-sized registers
20509 // and mixing the two types creates horrible code. This method optimizes
20510 // some of the transition sequences.
20511 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
20512 TargetLowering::DAGCombinerInfo &DCI,
20513 const X86Subtarget *Subtarget) {
20514 EVT VT = N->getValueType(0);
20515 if (!VT.is256BitVector())
20518 assert((N->getOpcode() == ISD::ANY_EXTEND ||
20519 N->getOpcode() == ISD::ZERO_EXTEND ||
20520 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
20522 SDValue Narrow = N->getOperand(0);
20523 EVT NarrowVT = Narrow->getValueType(0);
20524 if (!NarrowVT.is128BitVector())
20527 if (Narrow->getOpcode() != ISD::XOR &&
20528 Narrow->getOpcode() != ISD::AND &&
20529 Narrow->getOpcode() != ISD::OR)
20532 SDValue N0 = Narrow->getOperand(0);
20533 SDValue N1 = Narrow->getOperand(1);
20536 // The Left side has to be a trunc.
20537 if (N0.getOpcode() != ISD::TRUNCATE)
20540 // The type of the truncated inputs.
20541 EVT WideVT = N0->getOperand(0)->getValueType(0);
20545 // The right side has to be a 'trunc' or a constant vector.
20546 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
20547 ConstantSDNode *RHSConstSplat = nullptr;
20548 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
20549 RHSConstSplat = RHSBV->getConstantSplatNode();
20550 if (!RHSTrunc && !RHSConstSplat)
20553 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20555 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
20558 // Set N0 and N1 to hold the inputs to the new wide operation.
20559 N0 = N0->getOperand(0);
20560 if (RHSConstSplat) {
20561 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
20562 SDValue(RHSConstSplat, 0));
20563 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
20564 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
20565 } else if (RHSTrunc) {
20566 N1 = N1->getOperand(0);
20569 // Generate the wide operation.
20570 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
20571 unsigned Opcode = N->getOpcode();
20573 case ISD::ANY_EXTEND:
20575 case ISD::ZERO_EXTEND: {
20576 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
20577 APInt Mask = APInt::getAllOnesValue(InBits);
20578 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
20579 return DAG.getNode(ISD::AND, DL, VT,
20580 Op, DAG.getConstant(Mask, VT));
20582 case ISD::SIGN_EXTEND:
20583 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
20584 Op, DAG.getValueType(NarrowVT));
20586 llvm_unreachable("Unexpected opcode");
20590 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
20591 TargetLowering::DAGCombinerInfo &DCI,
20592 const X86Subtarget *Subtarget) {
20593 EVT VT = N->getValueType(0);
20594 if (DCI.isBeforeLegalizeOps())
20597 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
20601 // Create BEXTR instructions
20602 // BEXTR is ((X >> imm) & (2**size-1))
20603 if (VT == MVT::i32 || VT == MVT::i64) {
20604 SDValue N0 = N->getOperand(0);
20605 SDValue N1 = N->getOperand(1);
20608 // Check for BEXTR.
20609 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
20610 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
20611 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
20612 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
20613 if (MaskNode && ShiftNode) {
20614 uint64_t Mask = MaskNode->getZExtValue();
20615 uint64_t Shift = ShiftNode->getZExtValue();
20616 if (isMask_64(Mask)) {
20617 uint64_t MaskSize = CountPopulation_64(Mask);
20618 if (Shift + MaskSize <= VT.getSizeInBits())
20619 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
20620 DAG.getConstant(Shift | (MaskSize << 8), VT));
20628 // Want to form ANDNP nodes:
20629 // 1) In the hopes of then easily combining them with OR and AND nodes
20630 // to form PBLEND/PSIGN.
20631 // 2) To match ANDN packed intrinsics
20632 if (VT != MVT::v2i64 && VT != MVT::v4i64)
20635 SDValue N0 = N->getOperand(0);
20636 SDValue N1 = N->getOperand(1);
20639 // Check LHS for vnot
20640 if (N0.getOpcode() == ISD::XOR &&
20641 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
20642 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
20643 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
20645 // Check RHS for vnot
20646 if (N1.getOpcode() == ISD::XOR &&
20647 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
20648 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
20649 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
20654 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
20655 TargetLowering::DAGCombinerInfo &DCI,
20656 const X86Subtarget *Subtarget) {
20657 if (DCI.isBeforeLegalizeOps())
20660 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
20664 SDValue N0 = N->getOperand(0);
20665 SDValue N1 = N->getOperand(1);
20666 EVT VT = N->getValueType(0);
20668 // look for psign/blend
20669 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
20670 if (!Subtarget->hasSSSE3() ||
20671 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
20674 // Canonicalize pandn to RHS
20675 if (N0.getOpcode() == X86ISD::ANDNP)
20677 // or (and (m, y), (pandn m, x))
20678 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
20679 SDValue Mask = N1.getOperand(0);
20680 SDValue X = N1.getOperand(1);
20682 if (N0.getOperand(0) == Mask)
20683 Y = N0.getOperand(1);
20684 if (N0.getOperand(1) == Mask)
20685 Y = N0.getOperand(0);
20687 // Check to see if the mask appeared in both the AND and ANDNP and
20691 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
20692 // Look through mask bitcast.
20693 if (Mask.getOpcode() == ISD::BITCAST)
20694 Mask = Mask.getOperand(0);
20695 if (X.getOpcode() == ISD::BITCAST)
20696 X = X.getOperand(0);
20697 if (Y.getOpcode() == ISD::BITCAST)
20698 Y = Y.getOperand(0);
20700 EVT MaskVT = Mask.getValueType();
20702 // Validate that the Mask operand is a vector sra node.
20703 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
20704 // there is no psrai.b
20705 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
20706 unsigned SraAmt = ~0;
20707 if (Mask.getOpcode() == ISD::SRA) {
20708 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
20709 if (auto *AmtConst = AmtBV->getConstantSplatNode())
20710 SraAmt = AmtConst->getZExtValue();
20711 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
20712 SDValue SraC = Mask.getOperand(1);
20713 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
20715 if ((SraAmt + 1) != EltBits)
20720 // Now we know we at least have a plendvb with the mask val. See if
20721 // we can form a psignb/w/d.
20722 // psign = x.type == y.type == mask.type && y = sub(0, x);
20723 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
20724 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
20725 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
20726 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
20727 "Unsupported VT for PSIGN");
20728 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
20729 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
20731 // PBLENDVB only available on SSE 4.1
20732 if (!Subtarget->hasSSE41())
20735 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
20737 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
20738 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
20739 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
20740 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
20741 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
20745 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
20748 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
20749 MachineFunction &MF = DAG.getMachineFunction();
20750 bool OptForSize = MF.getFunction()->getAttributes().
20751 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
20753 // SHLD/SHRD instructions have lower register pressure, but on some
20754 // platforms they have higher latency than the equivalent
20755 // series of shifts/or that would otherwise be generated.
20756 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
20757 // have higher latencies and we are not optimizing for size.
20758 if (!OptForSize && Subtarget->isSHLDSlow())
20761 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
20763 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
20765 if (!N0.hasOneUse() || !N1.hasOneUse())
20768 SDValue ShAmt0 = N0.getOperand(1);
20769 if (ShAmt0.getValueType() != MVT::i8)
20771 SDValue ShAmt1 = N1.getOperand(1);
20772 if (ShAmt1.getValueType() != MVT::i8)
20774 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
20775 ShAmt0 = ShAmt0.getOperand(0);
20776 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
20777 ShAmt1 = ShAmt1.getOperand(0);
20780 unsigned Opc = X86ISD::SHLD;
20781 SDValue Op0 = N0.getOperand(0);
20782 SDValue Op1 = N1.getOperand(0);
20783 if (ShAmt0.getOpcode() == ISD::SUB) {
20784 Opc = X86ISD::SHRD;
20785 std::swap(Op0, Op1);
20786 std::swap(ShAmt0, ShAmt1);
20789 unsigned Bits = VT.getSizeInBits();
20790 if (ShAmt1.getOpcode() == ISD::SUB) {
20791 SDValue Sum = ShAmt1.getOperand(0);
20792 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
20793 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
20794 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
20795 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
20796 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
20797 return DAG.getNode(Opc, DL, VT,
20799 DAG.getNode(ISD::TRUNCATE, DL,
20802 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
20803 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
20805 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
20806 return DAG.getNode(Opc, DL, VT,
20807 N0.getOperand(0), N1.getOperand(0),
20808 DAG.getNode(ISD::TRUNCATE, DL,
20815 // Generate NEG and CMOV for integer abs.
20816 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
20817 EVT VT = N->getValueType(0);
20819 // Since X86 does not have CMOV for 8-bit integer, we don't convert
20820 // 8-bit integer abs to NEG and CMOV.
20821 if (VT.isInteger() && VT.getSizeInBits() == 8)
20824 SDValue N0 = N->getOperand(0);
20825 SDValue N1 = N->getOperand(1);
20828 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
20829 // and change it to SUB and CMOV.
20830 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
20831 N0.getOpcode() == ISD::ADD &&
20832 N0.getOperand(1) == N1 &&
20833 N1.getOpcode() == ISD::SRA &&
20834 N1.getOperand(0) == N0.getOperand(0))
20835 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
20836 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
20837 // Generate SUB & CMOV.
20838 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
20839 DAG.getConstant(0, VT), N0.getOperand(0));
20841 SDValue Ops[] = { N0.getOperand(0), Neg,
20842 DAG.getConstant(X86::COND_GE, MVT::i8),
20843 SDValue(Neg.getNode(), 1) };
20844 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
20849 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
20850 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
20851 TargetLowering::DAGCombinerInfo &DCI,
20852 const X86Subtarget *Subtarget) {
20853 if (DCI.isBeforeLegalizeOps())
20856 if (Subtarget->hasCMov()) {
20857 SDValue RV = performIntegerAbsCombine(N, DAG);
20865 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
20866 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
20867 TargetLowering::DAGCombinerInfo &DCI,
20868 const X86Subtarget *Subtarget) {
20869 LoadSDNode *Ld = cast<LoadSDNode>(N);
20870 EVT RegVT = Ld->getValueType(0);
20871 EVT MemVT = Ld->getMemoryVT();
20873 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20874 unsigned RegSz = RegVT.getSizeInBits();
20876 // On Sandybridge unaligned 256bit loads are inefficient.
20877 ISD::LoadExtType Ext = Ld->getExtensionType();
20878 unsigned Alignment = Ld->getAlignment();
20879 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
20880 if (RegVT.is256BitVector() && !Subtarget->hasInt256() &&
20881 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
20882 unsigned NumElems = RegVT.getVectorNumElements();
20886 SDValue Ptr = Ld->getBasePtr();
20887 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
20889 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
20891 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
20892 Ld->getPointerInfo(), Ld->isVolatile(),
20893 Ld->isNonTemporal(), Ld->isInvariant(),
20895 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
20896 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
20897 Ld->getPointerInfo(), Ld->isVolatile(),
20898 Ld->isNonTemporal(), Ld->isInvariant(),
20899 std::min(16U, Alignment));
20900 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
20902 Load2.getValue(1));
20904 SDValue NewVec = DAG.getUNDEF(RegVT);
20905 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
20906 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
20907 return DCI.CombineTo(N, NewVec, TF, true);
20910 // If this is a vector EXT Load then attempt to optimize it using a
20911 // shuffle. If SSSE3 is not available we may emit an illegal shuffle but the
20912 // expansion is still better than scalar code.
20913 // We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise we'll
20914 // emit a shuffle and a arithmetic shift.
20915 // TODO: It is possible to support ZExt by zeroing the undef values
20916 // during the shuffle phase or after the shuffle.
20917 if (RegVT.isVector() && RegVT.isInteger() && Subtarget->hasSSE2() &&
20918 (Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)) {
20919 assert(MemVT != RegVT && "Cannot extend to the same type");
20920 assert(MemVT.isVector() && "Must load a vector from memory");
20922 unsigned NumElems = RegVT.getVectorNumElements();
20923 unsigned MemSz = MemVT.getSizeInBits();
20924 assert(RegSz > MemSz && "Register size must be greater than the mem size");
20926 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256())
20929 // All sizes must be a power of two.
20930 if (!isPowerOf2_32(RegSz * MemSz * NumElems))
20933 // Attempt to load the original value using scalar loads.
20934 // Find the largest scalar type that divides the total loaded size.
20935 MVT SclrLoadTy = MVT::i8;
20936 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
20937 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
20938 MVT Tp = (MVT::SimpleValueType)tp;
20939 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
20944 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
20945 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
20947 SclrLoadTy = MVT::f64;
20949 // Calculate the number of scalar loads that we need to perform
20950 // in order to load our vector from memory.
20951 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
20952 if (Ext == ISD::SEXTLOAD && NumLoads > 1)
20955 unsigned loadRegZize = RegSz;
20956 if (Ext == ISD::SEXTLOAD && RegSz == 256)
20959 // Represent our vector as a sequence of elements which are the
20960 // largest scalar that we can load.
20961 EVT LoadUnitVecVT = EVT::getVectorVT(*DAG.getContext(), SclrLoadTy,
20962 loadRegZize/SclrLoadTy.getSizeInBits());
20964 // Represent the data using the same element type that is stored in
20965 // memory. In practice, we ''widen'' MemVT.
20967 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
20968 loadRegZize/MemVT.getScalarType().getSizeInBits());
20970 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
20971 "Invalid vector type");
20973 // We can't shuffle using an illegal type.
20974 if (!TLI.isTypeLegal(WideVecVT))
20977 SmallVector<SDValue, 8> Chains;
20978 SDValue Ptr = Ld->getBasePtr();
20979 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits()/8,
20980 TLI.getPointerTy());
20981 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
20983 for (unsigned i = 0; i < NumLoads; ++i) {
20984 // Perform a single load.
20985 SDValue ScalarLoad = DAG.getLoad(SclrLoadTy, dl, Ld->getChain(),
20986 Ptr, Ld->getPointerInfo(),
20987 Ld->isVolatile(), Ld->isNonTemporal(),
20988 Ld->isInvariant(), Ld->getAlignment());
20989 Chains.push_back(ScalarLoad.getValue(1));
20990 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
20991 // another round of DAGCombining.
20993 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
20995 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
20996 ScalarLoad, DAG.getIntPtrConstant(i));
20998 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
21001 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
21003 // Bitcast the loaded value to a vector of the original element type, in
21004 // the size of the target vector type.
21005 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
21006 unsigned SizeRatio = RegSz/MemSz;
21008 if (Ext == ISD::SEXTLOAD) {
21009 // If we have SSE4.1 we can directly emit a VSEXT node.
21010 if (Subtarget->hasSSE41()) {
21011 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
21012 return DCI.CombineTo(N, Sext, TF, true);
21015 // Otherwise we'll shuffle the small elements in the high bits of the
21016 // larger type and perform an arithmetic shift. If the shift is not legal
21017 // it's better to scalarize.
21018 if (!TLI.isOperationLegalOrCustom(ISD::SRA, RegVT))
21021 // Redistribute the loaded elements into the different locations.
21022 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
21023 for (unsigned i = 0; i != NumElems; ++i)
21024 ShuffleVec[i*SizeRatio + SizeRatio-1] = i;
21026 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
21027 DAG.getUNDEF(WideVecVT),
21030 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
21032 // Build the arithmetic shift.
21033 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
21034 MemVT.getVectorElementType().getSizeInBits();
21035 Shuff = DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
21036 DAG.getConstant(Amt, RegVT));
21038 return DCI.CombineTo(N, Shuff, TF, true);
21041 // Redistribute the loaded elements into the different locations.
21042 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
21043 for (unsigned i = 0; i != NumElems; ++i)
21044 ShuffleVec[i*SizeRatio] = i;
21046 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
21047 DAG.getUNDEF(WideVecVT),
21050 // Bitcast to the requested type.
21051 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
21052 // Replace the original load with the new sequence
21053 // and return the new chain.
21054 return DCI.CombineTo(N, Shuff, TF, true);
21060 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
21061 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
21062 const X86Subtarget *Subtarget) {
21063 StoreSDNode *St = cast<StoreSDNode>(N);
21064 EVT VT = St->getValue().getValueType();
21065 EVT StVT = St->getMemoryVT();
21067 SDValue StoredVal = St->getOperand(1);
21068 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21070 // If we are saving a concatenation of two XMM registers, perform two stores.
21071 // On Sandy Bridge, 256-bit memory operations are executed by two
21072 // 128-bit ports. However, on Haswell it is better to issue a single 256-bit
21073 // memory operation.
21074 unsigned Alignment = St->getAlignment();
21075 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
21076 if (VT.is256BitVector() && !Subtarget->hasInt256() &&
21077 StVT == VT && !IsAligned) {
21078 unsigned NumElems = VT.getVectorNumElements();
21082 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
21083 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
21085 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
21086 SDValue Ptr0 = St->getBasePtr();
21087 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
21089 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
21090 St->getPointerInfo(), St->isVolatile(),
21091 St->isNonTemporal(), Alignment);
21092 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
21093 St->getPointerInfo(), St->isVolatile(),
21094 St->isNonTemporal(),
21095 std::min(16U, Alignment));
21096 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
21099 // Optimize trunc store (of multiple scalars) to shuffle and store.
21100 // First, pack all of the elements in one place. Next, store to memory
21101 // in fewer chunks.
21102 if (St->isTruncatingStore() && VT.isVector()) {
21103 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21104 unsigned NumElems = VT.getVectorNumElements();
21105 assert(StVT != VT && "Cannot truncate to the same type");
21106 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
21107 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
21109 // From, To sizes and ElemCount must be pow of two
21110 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
21111 // We are going to use the original vector elt for storing.
21112 // Accumulated smaller vector elements must be a multiple of the store size.
21113 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
21115 unsigned SizeRatio = FromSz / ToSz;
21117 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
21119 // Create a type on which we perform the shuffle
21120 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
21121 StVT.getScalarType(), NumElems*SizeRatio);
21123 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
21125 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
21126 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
21127 for (unsigned i = 0; i != NumElems; ++i)
21128 ShuffleVec[i] = i * SizeRatio;
21130 // Can't shuffle using an illegal type.
21131 if (!TLI.isTypeLegal(WideVecVT))
21134 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
21135 DAG.getUNDEF(WideVecVT),
21137 // At this point all of the data is stored at the bottom of the
21138 // register. We now need to save it to mem.
21140 // Find the largest store unit
21141 MVT StoreType = MVT::i8;
21142 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
21143 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
21144 MVT Tp = (MVT::SimpleValueType)tp;
21145 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
21149 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
21150 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
21151 (64 <= NumElems * ToSz))
21152 StoreType = MVT::f64;
21154 // Bitcast the original vector into a vector of store-size units
21155 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
21156 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
21157 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
21158 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
21159 SmallVector<SDValue, 8> Chains;
21160 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
21161 TLI.getPointerTy());
21162 SDValue Ptr = St->getBasePtr();
21164 // Perform one or more big stores into memory.
21165 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
21166 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
21167 StoreType, ShuffWide,
21168 DAG.getIntPtrConstant(i));
21169 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
21170 St->getPointerInfo(), St->isVolatile(),
21171 St->isNonTemporal(), St->getAlignment());
21172 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
21173 Chains.push_back(Ch);
21176 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
21179 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
21180 // the FP state in cases where an emms may be missing.
21181 // A preferable solution to the general problem is to figure out the right
21182 // places to insert EMMS. This qualifies as a quick hack.
21184 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
21185 if (VT.getSizeInBits() != 64)
21188 const Function *F = DAG.getMachineFunction().getFunction();
21189 bool NoImplicitFloatOps = F->getAttributes().
21190 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
21191 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
21192 && Subtarget->hasSSE2();
21193 if ((VT.isVector() ||
21194 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
21195 isa<LoadSDNode>(St->getValue()) &&
21196 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
21197 St->getChain().hasOneUse() && !St->isVolatile()) {
21198 SDNode* LdVal = St->getValue().getNode();
21199 LoadSDNode *Ld = nullptr;
21200 int TokenFactorIndex = -1;
21201 SmallVector<SDValue, 8> Ops;
21202 SDNode* ChainVal = St->getChain().getNode();
21203 // Must be a store of a load. We currently handle two cases: the load
21204 // is a direct child, and it's under an intervening TokenFactor. It is
21205 // possible to dig deeper under nested TokenFactors.
21206 if (ChainVal == LdVal)
21207 Ld = cast<LoadSDNode>(St->getChain());
21208 else if (St->getValue().hasOneUse() &&
21209 ChainVal->getOpcode() == ISD::TokenFactor) {
21210 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
21211 if (ChainVal->getOperand(i).getNode() == LdVal) {
21212 TokenFactorIndex = i;
21213 Ld = cast<LoadSDNode>(St->getValue());
21215 Ops.push_back(ChainVal->getOperand(i));
21219 if (!Ld || !ISD::isNormalLoad(Ld))
21222 // If this is not the MMX case, i.e. we are just turning i64 load/store
21223 // into f64 load/store, avoid the transformation if there are multiple
21224 // uses of the loaded value.
21225 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
21230 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
21231 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
21233 if (Subtarget->is64Bit() || F64IsLegal) {
21234 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
21235 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
21236 Ld->getPointerInfo(), Ld->isVolatile(),
21237 Ld->isNonTemporal(), Ld->isInvariant(),
21238 Ld->getAlignment());
21239 SDValue NewChain = NewLd.getValue(1);
21240 if (TokenFactorIndex != -1) {
21241 Ops.push_back(NewChain);
21242 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
21244 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
21245 St->getPointerInfo(),
21246 St->isVolatile(), St->isNonTemporal(),
21247 St->getAlignment());
21250 // Otherwise, lower to two pairs of 32-bit loads / stores.
21251 SDValue LoAddr = Ld->getBasePtr();
21252 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
21253 DAG.getConstant(4, MVT::i32));
21255 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
21256 Ld->getPointerInfo(),
21257 Ld->isVolatile(), Ld->isNonTemporal(),
21258 Ld->isInvariant(), Ld->getAlignment());
21259 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
21260 Ld->getPointerInfo().getWithOffset(4),
21261 Ld->isVolatile(), Ld->isNonTemporal(),
21263 MinAlign(Ld->getAlignment(), 4));
21265 SDValue NewChain = LoLd.getValue(1);
21266 if (TokenFactorIndex != -1) {
21267 Ops.push_back(LoLd);
21268 Ops.push_back(HiLd);
21269 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
21272 LoAddr = St->getBasePtr();
21273 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
21274 DAG.getConstant(4, MVT::i32));
21276 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
21277 St->getPointerInfo(),
21278 St->isVolatile(), St->isNonTemporal(),
21279 St->getAlignment());
21280 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
21281 St->getPointerInfo().getWithOffset(4),
21283 St->isNonTemporal(),
21284 MinAlign(St->getAlignment(), 4));
21285 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
21290 /// isHorizontalBinOp - Return 'true' if this vector operation is "horizontal"
21291 /// and return the operands for the horizontal operation in LHS and RHS. A
21292 /// horizontal operation performs the binary operation on successive elements
21293 /// of its first operand, then on successive elements of its second operand,
21294 /// returning the resulting values in a vector. For example, if
21295 /// A = < float a0, float a1, float a2, float a3 >
21297 /// B = < float b0, float b1, float b2, float b3 >
21298 /// then the result of doing a horizontal operation on A and B is
21299 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
21300 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
21301 /// A horizontal-op B, for some already available A and B, and if so then LHS is
21302 /// set to A, RHS to B, and the routine returns 'true'.
21303 /// Note that the binary operation should have the property that if one of the
21304 /// operands is UNDEF then the result is UNDEF.
21305 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
21306 // Look for the following pattern: if
21307 // A = < float a0, float a1, float a2, float a3 >
21308 // B = < float b0, float b1, float b2, float b3 >
21310 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
21311 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
21312 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
21313 // which is A horizontal-op B.
21315 // At least one of the operands should be a vector shuffle.
21316 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
21317 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
21320 MVT VT = LHS.getSimpleValueType();
21322 assert((VT.is128BitVector() || VT.is256BitVector()) &&
21323 "Unsupported vector type for horizontal add/sub");
21325 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
21326 // operate independently on 128-bit lanes.
21327 unsigned NumElts = VT.getVectorNumElements();
21328 unsigned NumLanes = VT.getSizeInBits()/128;
21329 unsigned NumLaneElts = NumElts / NumLanes;
21330 assert((NumLaneElts % 2 == 0) &&
21331 "Vector type should have an even number of elements in each lane");
21332 unsigned HalfLaneElts = NumLaneElts/2;
21334 // View LHS in the form
21335 // LHS = VECTOR_SHUFFLE A, B, LMask
21336 // If LHS is not a shuffle then pretend it is the shuffle
21337 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
21338 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
21341 SmallVector<int, 16> LMask(NumElts);
21342 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
21343 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
21344 A = LHS.getOperand(0);
21345 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
21346 B = LHS.getOperand(1);
21347 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
21348 std::copy(Mask.begin(), Mask.end(), LMask.begin());
21350 if (LHS.getOpcode() != ISD::UNDEF)
21352 for (unsigned i = 0; i != NumElts; ++i)
21356 // Likewise, view RHS in the form
21357 // RHS = VECTOR_SHUFFLE C, D, RMask
21359 SmallVector<int, 16> RMask(NumElts);
21360 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
21361 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
21362 C = RHS.getOperand(0);
21363 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
21364 D = RHS.getOperand(1);
21365 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
21366 std::copy(Mask.begin(), Mask.end(), RMask.begin());
21368 if (RHS.getOpcode() != ISD::UNDEF)
21370 for (unsigned i = 0; i != NumElts; ++i)
21374 // Check that the shuffles are both shuffling the same vectors.
21375 if (!(A == C && B == D) && !(A == D && B == C))
21378 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
21379 if (!A.getNode() && !B.getNode())
21382 // If A and B occur in reverse order in RHS, then "swap" them (which means
21383 // rewriting the mask).
21385 CommuteVectorShuffleMask(RMask, NumElts);
21387 // At this point LHS and RHS are equivalent to
21388 // LHS = VECTOR_SHUFFLE A, B, LMask
21389 // RHS = VECTOR_SHUFFLE A, B, RMask
21390 // Check that the masks correspond to performing a horizontal operation.
21391 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
21392 for (unsigned i = 0; i != NumLaneElts; ++i) {
21393 int LIdx = LMask[i+l], RIdx = RMask[i+l];
21395 // Ignore any UNDEF components.
21396 if (LIdx < 0 || RIdx < 0 ||
21397 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
21398 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
21401 // Check that successive elements are being operated on. If not, this is
21402 // not a horizontal operation.
21403 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
21404 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
21405 if (!(LIdx == Index && RIdx == Index + 1) &&
21406 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
21411 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
21412 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
21416 /// PerformFADDCombine - Do target-specific dag combines on floating point adds.
21417 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
21418 const X86Subtarget *Subtarget) {
21419 EVT VT = N->getValueType(0);
21420 SDValue LHS = N->getOperand(0);
21421 SDValue RHS = N->getOperand(1);
21423 // Try to synthesize horizontal adds from adds of shuffles.
21424 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
21425 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
21426 isHorizontalBinOp(LHS, RHS, true))
21427 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
21431 /// PerformFSUBCombine - Do target-specific dag combines on floating point subs.
21432 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
21433 const X86Subtarget *Subtarget) {
21434 EVT VT = N->getValueType(0);
21435 SDValue LHS = N->getOperand(0);
21436 SDValue RHS = N->getOperand(1);
21438 // Try to synthesize horizontal subs from subs of shuffles.
21439 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
21440 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
21441 isHorizontalBinOp(LHS, RHS, false))
21442 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
21446 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
21447 /// X86ISD::FXOR nodes.
21448 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
21449 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
21450 // F[X]OR(0.0, x) -> x
21451 // F[X]OR(x, 0.0) -> x
21452 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
21453 if (C->getValueAPF().isPosZero())
21454 return N->getOperand(1);
21455 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
21456 if (C->getValueAPF().isPosZero())
21457 return N->getOperand(0);
21461 /// PerformFMinFMaxCombine - Do target-specific dag combines on X86ISD::FMIN and
21462 /// X86ISD::FMAX nodes.
21463 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
21464 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
21466 // Only perform optimizations if UnsafeMath is used.
21467 if (!DAG.getTarget().Options.UnsafeFPMath)
21470 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
21471 // into FMINC and FMAXC, which are Commutative operations.
21472 unsigned NewOp = 0;
21473 switch (N->getOpcode()) {
21474 default: llvm_unreachable("unknown opcode");
21475 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
21476 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
21479 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
21480 N->getOperand(0), N->getOperand(1));
21483 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
21484 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
21485 // FAND(0.0, x) -> 0.0
21486 // FAND(x, 0.0) -> 0.0
21487 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
21488 if (C->getValueAPF().isPosZero())
21489 return N->getOperand(0);
21490 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
21491 if (C->getValueAPF().isPosZero())
21492 return N->getOperand(1);
21496 /// PerformFANDNCombine - Do target-specific dag combines on X86ISD::FANDN nodes
21497 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
21498 // FANDN(x, 0.0) -> 0.0
21499 // FANDN(0.0, x) -> x
21500 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
21501 if (C->getValueAPF().isPosZero())
21502 return N->getOperand(1);
21503 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
21504 if (C->getValueAPF().isPosZero())
21505 return N->getOperand(1);
21509 static SDValue PerformBTCombine(SDNode *N,
21511 TargetLowering::DAGCombinerInfo &DCI) {
21512 // BT ignores high bits in the bit index operand.
21513 SDValue Op1 = N->getOperand(1);
21514 if (Op1.hasOneUse()) {
21515 unsigned BitWidth = Op1.getValueSizeInBits();
21516 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
21517 APInt KnownZero, KnownOne;
21518 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
21519 !DCI.isBeforeLegalizeOps());
21520 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21521 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
21522 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
21523 DCI.CommitTargetLoweringOpt(TLO);
21528 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
21529 SDValue Op = N->getOperand(0);
21530 if (Op.getOpcode() == ISD::BITCAST)
21531 Op = Op.getOperand(0);
21532 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
21533 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
21534 VT.getVectorElementType().getSizeInBits() ==
21535 OpVT.getVectorElementType().getSizeInBits()) {
21536 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
21541 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
21542 const X86Subtarget *Subtarget) {
21543 EVT VT = N->getValueType(0);
21544 if (!VT.isVector())
21547 SDValue N0 = N->getOperand(0);
21548 SDValue N1 = N->getOperand(1);
21549 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
21552 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
21553 // both SSE and AVX2 since there is no sign-extended shift right
21554 // operation on a vector with 64-bit elements.
21555 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
21556 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
21557 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
21558 N0.getOpcode() == ISD::SIGN_EXTEND)) {
21559 SDValue N00 = N0.getOperand(0);
21561 // EXTLOAD has a better solution on AVX2,
21562 // it may be replaced with X86ISD::VSEXT node.
21563 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
21564 if (!ISD::isNormalLoad(N00.getNode()))
21567 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
21568 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
21570 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
21576 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
21577 TargetLowering::DAGCombinerInfo &DCI,
21578 const X86Subtarget *Subtarget) {
21579 if (!DCI.isBeforeLegalizeOps())
21582 if (!Subtarget->hasFp256())
21585 EVT VT = N->getValueType(0);
21586 if (VT.isVector() && VT.getSizeInBits() == 256) {
21587 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
21595 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
21596 const X86Subtarget* Subtarget) {
21598 EVT VT = N->getValueType(0);
21600 // Let legalize expand this if it isn't a legal type yet.
21601 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
21604 EVT ScalarVT = VT.getScalarType();
21605 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
21606 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
21609 SDValue A = N->getOperand(0);
21610 SDValue B = N->getOperand(1);
21611 SDValue C = N->getOperand(2);
21613 bool NegA = (A.getOpcode() == ISD::FNEG);
21614 bool NegB = (B.getOpcode() == ISD::FNEG);
21615 bool NegC = (C.getOpcode() == ISD::FNEG);
21617 // Negative multiplication when NegA xor NegB
21618 bool NegMul = (NegA != NegB);
21620 A = A.getOperand(0);
21622 B = B.getOperand(0);
21624 C = C.getOperand(0);
21628 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
21630 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
21632 return DAG.getNode(Opcode, dl, VT, A, B, C);
21635 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
21636 TargetLowering::DAGCombinerInfo &DCI,
21637 const X86Subtarget *Subtarget) {
21638 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
21639 // (and (i32 x86isd::setcc_carry), 1)
21640 // This eliminates the zext. This transformation is necessary because
21641 // ISD::SETCC is always legalized to i8.
21643 SDValue N0 = N->getOperand(0);
21644 EVT VT = N->getValueType(0);
21646 if (N0.getOpcode() == ISD::AND &&
21648 N0.getOperand(0).hasOneUse()) {
21649 SDValue N00 = N0.getOperand(0);
21650 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
21651 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
21652 if (!C || C->getZExtValue() != 1)
21654 return DAG.getNode(ISD::AND, dl, VT,
21655 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
21656 N00.getOperand(0), N00.getOperand(1)),
21657 DAG.getConstant(1, VT));
21661 if (N0.getOpcode() == ISD::TRUNCATE &&
21663 N0.getOperand(0).hasOneUse()) {
21664 SDValue N00 = N0.getOperand(0);
21665 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
21666 return DAG.getNode(ISD::AND, dl, VT,
21667 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
21668 N00.getOperand(0), N00.getOperand(1)),
21669 DAG.getConstant(1, VT));
21672 if (VT.is256BitVector()) {
21673 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
21681 // Optimize x == -y --> x+y == 0
21682 // x != -y --> x+y != 0
21683 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
21684 const X86Subtarget* Subtarget) {
21685 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
21686 SDValue LHS = N->getOperand(0);
21687 SDValue RHS = N->getOperand(1);
21688 EVT VT = N->getValueType(0);
21691 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
21692 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
21693 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
21694 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
21695 LHS.getValueType(), RHS, LHS.getOperand(1));
21696 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
21697 addV, DAG.getConstant(0, addV.getValueType()), CC);
21699 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
21700 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
21701 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
21702 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
21703 RHS.getValueType(), LHS, RHS.getOperand(1));
21704 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
21705 addV, DAG.getConstant(0, addV.getValueType()), CC);
21708 if (VT.getScalarType() == MVT::i1) {
21709 bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
21710 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
21711 bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
21712 if (!IsSEXT0 && !IsVZero0)
21714 bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
21715 (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
21716 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
21718 if (!IsSEXT1 && !IsVZero1)
21721 if (IsSEXT0 && IsVZero1) {
21722 assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
21723 if (CC == ISD::SETEQ)
21724 return DAG.getNOT(DL, LHS.getOperand(0), VT);
21725 return LHS.getOperand(0);
21727 if (IsSEXT1 && IsVZero0) {
21728 assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
21729 if (CC == ISD::SETEQ)
21730 return DAG.getNOT(DL, RHS.getOperand(0), VT);
21731 return RHS.getOperand(0);
21738 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
21739 const X86Subtarget *Subtarget) {
21741 MVT VT = N->getOperand(1)->getSimpleValueType(0);
21742 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
21743 "X86insertps is only defined for v4x32");
21745 SDValue Ld = N->getOperand(1);
21746 if (MayFoldLoad(Ld)) {
21747 // Extract the countS bits from the immediate so we can get the proper
21748 // address when narrowing the vector load to a specific element.
21749 // When the second source op is a memory address, interps doesn't use
21750 // countS and just gets an f32 from that address.
21751 unsigned DestIndex =
21752 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
21753 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
21757 // Create this as a scalar to vector to match the instruction pattern.
21758 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
21759 // countS bits are ignored when loading from memory on insertps, which
21760 // means we don't need to explicitly set them to 0.
21761 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
21762 LoadScalarToVector, N->getOperand(2));
21765 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
21766 // as "sbb reg,reg", since it can be extended without zext and produces
21767 // an all-ones bit which is more useful than 0/1 in some cases.
21768 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
21771 return DAG.getNode(ISD::AND, DL, VT,
21772 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
21773 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
21774 DAG.getConstant(1, VT));
21775 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
21776 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
21777 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
21778 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS));
21781 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
21782 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
21783 TargetLowering::DAGCombinerInfo &DCI,
21784 const X86Subtarget *Subtarget) {
21786 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
21787 SDValue EFLAGS = N->getOperand(1);
21789 if (CC == X86::COND_A) {
21790 // Try to convert COND_A into COND_B in an attempt to facilitate
21791 // materializing "setb reg".
21793 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
21794 // cannot take an immediate as its first operand.
21796 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
21797 EFLAGS.getValueType().isInteger() &&
21798 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
21799 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
21800 EFLAGS.getNode()->getVTList(),
21801 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
21802 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
21803 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
21807 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
21808 // a zext and produces an all-ones bit which is more useful than 0/1 in some
21810 if (CC == X86::COND_B)
21811 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
21815 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
21816 if (Flags.getNode()) {
21817 SDValue Cond = DAG.getConstant(CC, MVT::i8);
21818 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
21824 // Optimize branch condition evaluation.
21826 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
21827 TargetLowering::DAGCombinerInfo &DCI,
21828 const X86Subtarget *Subtarget) {
21830 SDValue Chain = N->getOperand(0);
21831 SDValue Dest = N->getOperand(1);
21832 SDValue EFLAGS = N->getOperand(3);
21833 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
21837 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
21838 if (Flags.getNode()) {
21839 SDValue Cond = DAG.getConstant(CC, MVT::i8);
21840 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
21847 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
21848 SelectionDAG &DAG) {
21849 // Take advantage of vector comparisons producing 0 or -1 in each lane to
21850 // optimize away operation when it's from a constant.
21852 // The general transformation is:
21853 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
21854 // AND(VECTOR_CMP(x,y), constant2)
21855 // constant2 = UNARYOP(constant)
21857 // Early exit if this isn't a vector operation, the operand of the
21858 // unary operation isn't a bitwise AND, or if the sizes of the operations
21859 // aren't the same.
21860 EVT VT = N->getValueType(0);
21861 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
21862 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
21863 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
21866 // Now check that the other operand of the AND is a constant splat. We could
21867 // make the transformation for non-constant splats as well, but it's unclear
21868 // that would be a benefit as it would not eliminate any operations, just
21869 // perform one more step in scalar code before moving to the vector unit.
21870 if (BuildVectorSDNode *BV =
21871 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
21872 // Bail out if the vector isn't a constant splat.
21873 if (!BV->getConstantSplatNode())
21876 // Everything checks out. Build up the new and improved node.
21878 EVT IntVT = BV->getValueType(0);
21879 // Create a new constant of the appropriate type for the transformed
21881 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
21882 // The AND node needs bitcasts to/from an integer vector type around it.
21883 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
21884 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
21885 N->getOperand(0)->getOperand(0), MaskConst);
21886 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
21893 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
21894 const X86TargetLowering *XTLI) {
21895 // First try to optimize away the conversion entirely when it's
21896 // conditionally from a constant. Vectors only.
21897 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
21898 if (Res != SDValue())
21901 // Now move on to more general possibilities.
21902 SDValue Op0 = N->getOperand(0);
21903 EVT InVT = Op0->getValueType(0);
21905 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
21906 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
21908 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
21909 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
21910 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
21913 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
21914 // a 32-bit target where SSE doesn't support i64->FP operations.
21915 if (Op0.getOpcode() == ISD::LOAD) {
21916 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
21917 EVT VT = Ld->getValueType(0);
21918 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
21919 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
21920 !XTLI->getSubtarget()->is64Bit() &&
21922 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
21923 Ld->getChain(), Op0, DAG);
21924 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
21931 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
21932 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
21933 X86TargetLowering::DAGCombinerInfo &DCI) {
21934 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
21935 // the result is either zero or one (depending on the input carry bit).
21936 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
21937 if (X86::isZeroNode(N->getOperand(0)) &&
21938 X86::isZeroNode(N->getOperand(1)) &&
21939 // We don't have a good way to replace an EFLAGS use, so only do this when
21941 SDValue(N, 1).use_empty()) {
21943 EVT VT = N->getValueType(0);
21944 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
21945 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
21946 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
21947 DAG.getConstant(X86::COND_B,MVT::i8),
21949 DAG.getConstant(1, VT));
21950 return DCI.CombineTo(N, Res1, CarryOut);
21956 // fold (add Y, (sete X, 0)) -> adc 0, Y
21957 // (add Y, (setne X, 0)) -> sbb -1, Y
21958 // (sub (sete X, 0), Y) -> sbb 0, Y
21959 // (sub (setne X, 0), Y) -> adc -1, Y
21960 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
21963 // Look through ZExts.
21964 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
21965 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
21968 SDValue SetCC = Ext.getOperand(0);
21969 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
21972 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
21973 if (CC != X86::COND_E && CC != X86::COND_NE)
21976 SDValue Cmp = SetCC.getOperand(1);
21977 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
21978 !X86::isZeroNode(Cmp.getOperand(1)) ||
21979 !Cmp.getOperand(0).getValueType().isInteger())
21982 SDValue CmpOp0 = Cmp.getOperand(0);
21983 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
21984 DAG.getConstant(1, CmpOp0.getValueType()));
21986 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
21987 if (CC == X86::COND_NE)
21988 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
21989 DL, OtherVal.getValueType(), OtherVal,
21990 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
21991 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
21992 DL, OtherVal.getValueType(), OtherVal,
21993 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
21996 /// PerformADDCombine - Do target-specific dag combines on integer adds.
21997 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
21998 const X86Subtarget *Subtarget) {
21999 EVT VT = N->getValueType(0);
22000 SDValue Op0 = N->getOperand(0);
22001 SDValue Op1 = N->getOperand(1);
22003 // Try to synthesize horizontal adds from adds of shuffles.
22004 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
22005 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
22006 isHorizontalBinOp(Op0, Op1, true))
22007 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
22009 return OptimizeConditionalInDecrement(N, DAG);
22012 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
22013 const X86Subtarget *Subtarget) {
22014 SDValue Op0 = N->getOperand(0);
22015 SDValue Op1 = N->getOperand(1);
22017 // X86 can't encode an immediate LHS of a sub. See if we can push the
22018 // negation into a preceding instruction.
22019 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
22020 // If the RHS of the sub is a XOR with one use and a constant, invert the
22021 // immediate. Then add one to the LHS of the sub so we can turn
22022 // X-Y -> X+~Y+1, saving one register.
22023 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
22024 isa<ConstantSDNode>(Op1.getOperand(1))) {
22025 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
22026 EVT VT = Op0.getValueType();
22027 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
22029 DAG.getConstant(~XorC, VT));
22030 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
22031 DAG.getConstant(C->getAPIntValue()+1, VT));
22035 // Try to synthesize horizontal adds from adds of shuffles.
22036 EVT VT = N->getValueType(0);
22037 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
22038 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
22039 isHorizontalBinOp(Op0, Op1, true))
22040 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
22042 return OptimizeConditionalInDecrement(N, DAG);
22045 /// performVZEXTCombine - Performs build vector combines
22046 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
22047 TargetLowering::DAGCombinerInfo &DCI,
22048 const X86Subtarget *Subtarget) {
22049 // (vzext (bitcast (vzext (x)) -> (vzext x)
22050 SDValue In = N->getOperand(0);
22051 while (In.getOpcode() == ISD::BITCAST)
22052 In = In.getOperand(0);
22054 if (In.getOpcode() != X86ISD::VZEXT)
22057 return DAG.getNode(X86ISD::VZEXT, SDLoc(N), N->getValueType(0),
22061 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
22062 DAGCombinerInfo &DCI) const {
22063 SelectionDAG &DAG = DCI.DAG;
22064 switch (N->getOpcode()) {
22066 case ISD::EXTRACT_VECTOR_ELT:
22067 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
22069 case ISD::SELECT: return PerformSELECTCombine(N, DAG, DCI, Subtarget);
22070 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
22071 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
22072 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
22073 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
22074 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
22077 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
22078 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
22079 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
22080 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
22081 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
22082 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
22083 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
22084 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
22085 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
22087 case X86ISD::FOR: return PerformFORCombine(N, DAG);
22089 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
22090 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
22091 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
22092 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
22093 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
22094 case ISD::ANY_EXTEND:
22095 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
22096 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
22097 case ISD::SIGN_EXTEND_INREG:
22098 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
22099 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
22100 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
22101 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
22102 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
22103 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
22104 case X86ISD::SHUFP: // Handle all target specific shuffles
22105 case X86ISD::PALIGNR:
22106 case X86ISD::UNPCKH:
22107 case X86ISD::UNPCKL:
22108 case X86ISD::MOVHLPS:
22109 case X86ISD::MOVLHPS:
22110 case X86ISD::PSHUFD:
22111 case X86ISD::PSHUFHW:
22112 case X86ISD::PSHUFLW:
22113 case X86ISD::MOVSS:
22114 case X86ISD::MOVSD:
22115 case X86ISD::VPERMILP:
22116 case X86ISD::VPERM2X128:
22117 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
22118 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
22119 case ISD::INTRINSIC_WO_CHAIN:
22120 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
22121 case X86ISD::INSERTPS:
22122 return PerformINSERTPSCombine(N, DAG, Subtarget);
22123 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
22129 /// isTypeDesirableForOp - Return true if the target has native support for
22130 /// the specified value type and it is 'desirable' to use the type for the
22131 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
22132 /// instruction encodings are longer and some i16 instructions are slow.
22133 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
22134 if (!isTypeLegal(VT))
22136 if (VT != MVT::i16)
22143 case ISD::SIGN_EXTEND:
22144 case ISD::ZERO_EXTEND:
22145 case ISD::ANY_EXTEND:
22158 /// IsDesirableToPromoteOp - This method query the target whether it is
22159 /// beneficial for dag combiner to promote the specified node. If true, it
22160 /// should return the desired promotion type by reference.
22161 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
22162 EVT VT = Op.getValueType();
22163 if (VT != MVT::i16)
22166 bool Promote = false;
22167 bool Commute = false;
22168 switch (Op.getOpcode()) {
22171 LoadSDNode *LD = cast<LoadSDNode>(Op);
22172 // If the non-extending load has a single use and it's not live out, then it
22173 // might be folded.
22174 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
22175 Op.hasOneUse()*/) {
22176 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
22177 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
22178 // The only case where we'd want to promote LOAD (rather then it being
22179 // promoted as an operand is when it's only use is liveout.
22180 if (UI->getOpcode() != ISD::CopyToReg)
22187 case ISD::SIGN_EXTEND:
22188 case ISD::ZERO_EXTEND:
22189 case ISD::ANY_EXTEND:
22194 SDValue N0 = Op.getOperand(0);
22195 // Look out for (store (shl (load), x)).
22196 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
22209 SDValue N0 = Op.getOperand(0);
22210 SDValue N1 = Op.getOperand(1);
22211 if (!Commute && MayFoldLoad(N1))
22213 // Avoid disabling potential load folding opportunities.
22214 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
22216 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
22226 //===----------------------------------------------------------------------===//
22227 // X86 Inline Assembly Support
22228 //===----------------------------------------------------------------------===//
22231 // Helper to match a string separated by whitespace.
22232 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
22233 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
22235 for (unsigned i = 0, e = args.size(); i != e; ++i) {
22236 StringRef piece(*args[i]);
22237 if (!s.startswith(piece)) // Check if the piece matches.
22240 s = s.substr(piece.size());
22241 StringRef::size_type pos = s.find_first_not_of(" \t");
22242 if (pos == 0) // We matched a prefix.
22250 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
22253 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
22255 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
22256 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
22257 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
22258 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
22260 if (AsmPieces.size() == 3)
22262 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
22269 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
22270 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
22272 std::string AsmStr = IA->getAsmString();
22274 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
22275 if (!Ty || Ty->getBitWidth() % 16 != 0)
22278 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
22279 SmallVector<StringRef, 4> AsmPieces;
22280 SplitString(AsmStr, AsmPieces, ";\n");
22282 switch (AsmPieces.size()) {
22283 default: return false;
22285 // FIXME: this should verify that we are targeting a 486 or better. If not,
22286 // we will turn this bswap into something that will be lowered to logical
22287 // ops instead of emitting the bswap asm. For now, we don't support 486 or
22288 // lower so don't worry about this.
22290 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
22291 matchAsm(AsmPieces[0], "bswapl", "$0") ||
22292 matchAsm(AsmPieces[0], "bswapq", "$0") ||
22293 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
22294 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
22295 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
22296 // No need to check constraints, nothing other than the equivalent of
22297 // "=r,0" would be valid here.
22298 return IntrinsicLowering::LowerToByteSwap(CI);
22301 // rorw $$8, ${0:w} --> llvm.bswap.i16
22302 if (CI->getType()->isIntegerTy(16) &&
22303 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
22304 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
22305 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
22307 const std::string &ConstraintsStr = IA->getConstraintString();
22308 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
22309 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
22310 if (clobbersFlagRegisters(AsmPieces))
22311 return IntrinsicLowering::LowerToByteSwap(CI);
22315 if (CI->getType()->isIntegerTy(32) &&
22316 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
22317 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
22318 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
22319 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
22321 const std::string &ConstraintsStr = IA->getConstraintString();
22322 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
22323 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
22324 if (clobbersFlagRegisters(AsmPieces))
22325 return IntrinsicLowering::LowerToByteSwap(CI);
22328 if (CI->getType()->isIntegerTy(64)) {
22329 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
22330 if (Constraints.size() >= 2 &&
22331 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
22332 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
22333 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
22334 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
22335 matchAsm(AsmPieces[1], "bswap", "%edx") &&
22336 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
22337 return IntrinsicLowering::LowerToByteSwap(CI);
22345 /// getConstraintType - Given a constraint letter, return the type of
22346 /// constraint it is for this target.
22347 X86TargetLowering::ConstraintType
22348 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
22349 if (Constraint.size() == 1) {
22350 switch (Constraint[0]) {
22361 return C_RegisterClass;
22385 return TargetLowering::getConstraintType(Constraint);
22388 /// Examine constraint type and operand type and determine a weight value.
22389 /// This object must already have been set up with the operand type
22390 /// and the current alternative constraint selected.
22391 TargetLowering::ConstraintWeight
22392 X86TargetLowering::getSingleConstraintMatchWeight(
22393 AsmOperandInfo &info, const char *constraint) const {
22394 ConstraintWeight weight = CW_Invalid;
22395 Value *CallOperandVal = info.CallOperandVal;
22396 // If we don't have a value, we can't do a match,
22397 // but allow it at the lowest weight.
22398 if (!CallOperandVal)
22400 Type *type = CallOperandVal->getType();
22401 // Look at the constraint type.
22402 switch (*constraint) {
22404 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
22415 if (CallOperandVal->getType()->isIntegerTy())
22416 weight = CW_SpecificReg;
22421 if (type->isFloatingPointTy())
22422 weight = CW_SpecificReg;
22425 if (type->isX86_MMXTy() && Subtarget->hasMMX())
22426 weight = CW_SpecificReg;
22430 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
22431 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
22432 weight = CW_Register;
22435 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
22436 if (C->getZExtValue() <= 31)
22437 weight = CW_Constant;
22441 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
22442 if (C->getZExtValue() <= 63)
22443 weight = CW_Constant;
22447 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
22448 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
22449 weight = CW_Constant;
22453 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
22454 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
22455 weight = CW_Constant;
22459 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
22460 if (C->getZExtValue() <= 3)
22461 weight = CW_Constant;
22465 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
22466 if (C->getZExtValue() <= 0xff)
22467 weight = CW_Constant;
22472 if (dyn_cast<ConstantFP>(CallOperandVal)) {
22473 weight = CW_Constant;
22477 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
22478 if ((C->getSExtValue() >= -0x80000000LL) &&
22479 (C->getSExtValue() <= 0x7fffffffLL))
22480 weight = CW_Constant;
22484 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
22485 if (C->getZExtValue() <= 0xffffffff)
22486 weight = CW_Constant;
22493 /// LowerXConstraint - try to replace an X constraint, which matches anything,
22494 /// with another that has more specific requirements based on the type of the
22495 /// corresponding operand.
22496 const char *X86TargetLowering::
22497 LowerXConstraint(EVT ConstraintVT) const {
22498 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
22499 // 'f' like normal targets.
22500 if (ConstraintVT.isFloatingPoint()) {
22501 if (Subtarget->hasSSE2())
22503 if (Subtarget->hasSSE1())
22507 return TargetLowering::LowerXConstraint(ConstraintVT);
22510 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
22511 /// vector. If it is invalid, don't add anything to Ops.
22512 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
22513 std::string &Constraint,
22514 std::vector<SDValue>&Ops,
22515 SelectionDAG &DAG) const {
22518 // Only support length 1 constraints for now.
22519 if (Constraint.length() > 1) return;
22521 char ConstraintLetter = Constraint[0];
22522 switch (ConstraintLetter) {
22525 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
22526 if (C->getZExtValue() <= 31) {
22527 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
22533 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
22534 if (C->getZExtValue() <= 63) {
22535 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
22541 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
22542 if (isInt<8>(C->getSExtValue())) {
22543 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
22549 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
22550 if (C->getZExtValue() <= 255) {
22551 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
22557 // 32-bit signed value
22558 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
22559 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
22560 C->getSExtValue())) {
22561 // Widen to 64 bits here to get it sign extended.
22562 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
22565 // FIXME gcc accepts some relocatable values here too, but only in certain
22566 // memory models; it's complicated.
22571 // 32-bit unsigned value
22572 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
22573 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
22574 C->getZExtValue())) {
22575 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
22579 // FIXME gcc accepts some relocatable values here too, but only in certain
22580 // memory models; it's complicated.
22584 // Literal immediates are always ok.
22585 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
22586 // Widen to 64 bits here to get it sign extended.
22587 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
22591 // In any sort of PIC mode addresses need to be computed at runtime by
22592 // adding in a register or some sort of table lookup. These can't
22593 // be used as immediates.
22594 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
22597 // If we are in non-pic codegen mode, we allow the address of a global (with
22598 // an optional displacement) to be used with 'i'.
22599 GlobalAddressSDNode *GA = nullptr;
22600 int64_t Offset = 0;
22602 // Match either (GA), (GA+C), (GA+C1+C2), etc.
22604 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
22605 Offset += GA->getOffset();
22607 } else if (Op.getOpcode() == ISD::ADD) {
22608 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
22609 Offset += C->getZExtValue();
22610 Op = Op.getOperand(0);
22613 } else if (Op.getOpcode() == ISD::SUB) {
22614 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
22615 Offset += -C->getZExtValue();
22616 Op = Op.getOperand(0);
22621 // Otherwise, this isn't something we can handle, reject it.
22625 const GlobalValue *GV = GA->getGlobal();
22626 // If we require an extra load to get this address, as in PIC mode, we
22627 // can't accept it.
22628 if (isGlobalStubReference(
22629 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
22632 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
22633 GA->getValueType(0), Offset);
22638 if (Result.getNode()) {
22639 Ops.push_back(Result);
22642 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
22645 std::pair<unsigned, const TargetRegisterClass*>
22646 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
22648 // First, see if this is a constraint that directly corresponds to an LLVM
22650 if (Constraint.size() == 1) {
22651 // GCC Constraint Letters
22652 switch (Constraint[0]) {
22654 // TODO: Slight differences here in allocation order and leaving
22655 // RIP in the class. Do they matter any more here than they do
22656 // in the normal allocation?
22657 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
22658 if (Subtarget->is64Bit()) {
22659 if (VT == MVT::i32 || VT == MVT::f32)
22660 return std::make_pair(0U, &X86::GR32RegClass);
22661 if (VT == MVT::i16)
22662 return std::make_pair(0U, &X86::GR16RegClass);
22663 if (VT == MVT::i8 || VT == MVT::i1)
22664 return std::make_pair(0U, &X86::GR8RegClass);
22665 if (VT == MVT::i64 || VT == MVT::f64)
22666 return std::make_pair(0U, &X86::GR64RegClass);
22669 // 32-bit fallthrough
22670 case 'Q': // Q_REGS
22671 if (VT == MVT::i32 || VT == MVT::f32)
22672 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
22673 if (VT == MVT::i16)
22674 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
22675 if (VT == MVT::i8 || VT == MVT::i1)
22676 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
22677 if (VT == MVT::i64)
22678 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
22680 case 'r': // GENERAL_REGS
22681 case 'l': // INDEX_REGS
22682 if (VT == MVT::i8 || VT == MVT::i1)
22683 return std::make_pair(0U, &X86::GR8RegClass);
22684 if (VT == MVT::i16)
22685 return std::make_pair(0U, &X86::GR16RegClass);
22686 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
22687 return std::make_pair(0U, &X86::GR32RegClass);
22688 return std::make_pair(0U, &X86::GR64RegClass);
22689 case 'R': // LEGACY_REGS
22690 if (VT == MVT::i8 || VT == MVT::i1)
22691 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
22692 if (VT == MVT::i16)
22693 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
22694 if (VT == MVT::i32 || !Subtarget->is64Bit())
22695 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
22696 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
22697 case 'f': // FP Stack registers.
22698 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
22699 // value to the correct fpstack register class.
22700 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
22701 return std::make_pair(0U, &X86::RFP32RegClass);
22702 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
22703 return std::make_pair(0U, &X86::RFP64RegClass);
22704 return std::make_pair(0U, &X86::RFP80RegClass);
22705 case 'y': // MMX_REGS if MMX allowed.
22706 if (!Subtarget->hasMMX()) break;
22707 return std::make_pair(0U, &X86::VR64RegClass);
22708 case 'Y': // SSE_REGS if SSE2 allowed
22709 if (!Subtarget->hasSSE2()) break;
22711 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
22712 if (!Subtarget->hasSSE1()) break;
22714 switch (VT.SimpleTy) {
22716 // Scalar SSE types.
22719 return std::make_pair(0U, &X86::FR32RegClass);
22722 return std::make_pair(0U, &X86::FR64RegClass);
22730 return std::make_pair(0U, &X86::VR128RegClass);
22738 return std::make_pair(0U, &X86::VR256RegClass);
22743 return std::make_pair(0U, &X86::VR512RegClass);
22749 // Use the default implementation in TargetLowering to convert the register
22750 // constraint into a member of a register class.
22751 std::pair<unsigned, const TargetRegisterClass*> Res;
22752 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
22754 // Not found as a standard register?
22756 // Map st(0) -> st(7) -> ST0
22757 if (Constraint.size() == 7 && Constraint[0] == '{' &&
22758 tolower(Constraint[1]) == 's' &&
22759 tolower(Constraint[2]) == 't' &&
22760 Constraint[3] == '(' &&
22761 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
22762 Constraint[5] == ')' &&
22763 Constraint[6] == '}') {
22765 Res.first = X86::ST0+Constraint[4]-'0';
22766 Res.second = &X86::RFP80RegClass;
22770 // GCC allows "st(0)" to be called just plain "st".
22771 if (StringRef("{st}").equals_lower(Constraint)) {
22772 Res.first = X86::ST0;
22773 Res.second = &X86::RFP80RegClass;
22778 if (StringRef("{flags}").equals_lower(Constraint)) {
22779 Res.first = X86::EFLAGS;
22780 Res.second = &X86::CCRRegClass;
22784 // 'A' means EAX + EDX.
22785 if (Constraint == "A") {
22786 Res.first = X86::EAX;
22787 Res.second = &X86::GR32_ADRegClass;
22793 // Otherwise, check to see if this is a register class of the wrong value
22794 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
22795 // turn into {ax},{dx}.
22796 if (Res.second->hasType(VT))
22797 return Res; // Correct type already, nothing to do.
22799 // All of the single-register GCC register classes map their values onto
22800 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
22801 // really want an 8-bit or 32-bit register, map to the appropriate register
22802 // class and return the appropriate register.
22803 if (Res.second == &X86::GR16RegClass) {
22804 if (VT == MVT::i8 || VT == MVT::i1) {
22805 unsigned DestReg = 0;
22806 switch (Res.first) {
22808 case X86::AX: DestReg = X86::AL; break;
22809 case X86::DX: DestReg = X86::DL; break;
22810 case X86::CX: DestReg = X86::CL; break;
22811 case X86::BX: DestReg = X86::BL; break;
22814 Res.first = DestReg;
22815 Res.second = &X86::GR8RegClass;
22817 } else if (VT == MVT::i32 || VT == MVT::f32) {
22818 unsigned DestReg = 0;
22819 switch (Res.first) {
22821 case X86::AX: DestReg = X86::EAX; break;
22822 case X86::DX: DestReg = X86::EDX; break;
22823 case X86::CX: DestReg = X86::ECX; break;
22824 case X86::BX: DestReg = X86::EBX; break;
22825 case X86::SI: DestReg = X86::ESI; break;
22826 case X86::DI: DestReg = X86::EDI; break;
22827 case X86::BP: DestReg = X86::EBP; break;
22828 case X86::SP: DestReg = X86::ESP; break;
22831 Res.first = DestReg;
22832 Res.second = &X86::GR32RegClass;
22834 } else if (VT == MVT::i64 || VT == MVT::f64) {
22835 unsigned DestReg = 0;
22836 switch (Res.first) {
22838 case X86::AX: DestReg = X86::RAX; break;
22839 case X86::DX: DestReg = X86::RDX; break;
22840 case X86::CX: DestReg = X86::RCX; break;
22841 case X86::BX: DestReg = X86::RBX; break;
22842 case X86::SI: DestReg = X86::RSI; break;
22843 case X86::DI: DestReg = X86::RDI; break;
22844 case X86::BP: DestReg = X86::RBP; break;
22845 case X86::SP: DestReg = X86::RSP; break;
22848 Res.first = DestReg;
22849 Res.second = &X86::GR64RegClass;
22852 } else if (Res.second == &X86::FR32RegClass ||
22853 Res.second == &X86::FR64RegClass ||
22854 Res.second == &X86::VR128RegClass ||
22855 Res.second == &X86::VR256RegClass ||
22856 Res.second == &X86::FR32XRegClass ||
22857 Res.second == &X86::FR64XRegClass ||
22858 Res.second == &X86::VR128XRegClass ||
22859 Res.second == &X86::VR256XRegClass ||
22860 Res.second == &X86::VR512RegClass) {
22861 // Handle references to XMM physical registers that got mapped into the
22862 // wrong class. This can happen with constraints like {xmm0} where the
22863 // target independent register mapper will just pick the first match it can
22864 // find, ignoring the required type.
22866 if (VT == MVT::f32 || VT == MVT::i32)
22867 Res.second = &X86::FR32RegClass;
22868 else if (VT == MVT::f64 || VT == MVT::i64)
22869 Res.second = &X86::FR64RegClass;
22870 else if (X86::VR128RegClass.hasType(VT))
22871 Res.second = &X86::VR128RegClass;
22872 else if (X86::VR256RegClass.hasType(VT))
22873 Res.second = &X86::VR256RegClass;
22874 else if (X86::VR512RegClass.hasType(VT))
22875 Res.second = &X86::VR512RegClass;
22881 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
22883 // Scaling factors are not free at all.
22884 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
22885 // will take 2 allocations in the out of order engine instead of 1
22886 // for plain addressing mode, i.e. inst (reg1).
22888 // vaddps (%rsi,%drx), %ymm0, %ymm1
22889 // Requires two allocations (one for the load, one for the computation)
22891 // vaddps (%rsi), %ymm0, %ymm1
22892 // Requires just 1 allocation, i.e., freeing allocations for other operations
22893 // and having less micro operations to execute.
22895 // For some X86 architectures, this is even worse because for instance for
22896 // stores, the complex addressing mode forces the instruction to use the
22897 // "load" ports instead of the dedicated "store" port.
22898 // E.g., on Haswell:
22899 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
22900 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
22901 if (isLegalAddressingMode(AM, Ty))
22902 // Scale represents reg2 * scale, thus account for 1
22903 // as soon as we use a second register.
22904 return AM.Scale != 0;
22908 bool X86TargetLowering::isTargetFTOL() const {
22909 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();